Top 30 Foundation Model Developers 2027

This report forms part of The Economy Rankings’ Artificial Intelligence & Data portfolio. The evaluated unit is the organization responsible for original training and continuing maintenance of a general-purpose foundation-model family.

Foundation model development is a different business from building an application around a model. The developer determines the architecture, training process and release conditions of a reusable system. This ranking begins at that upstream responsibility: original training must be documented before a familiar assistant or creative interface becomes eligible.

The market includes language models, multimodal understanding systems and models that generate images, video or audio. Those modalities are within the established series scope, but they do not share a single meaningful performance scale. The comparison therefore considers capability within the model program alongside evaluation, governance, usability and transparency.

A maintained model family matters more than an isolated announcement. Buyers and developers need identifiable artifacts or working access, enough documentation to understand limitations and a responsible organization that continues to operate the program. A laboratory that has only promised a model, a company reselling another developer’s endpoint and an application that only fine-tunes external weights do not satisfy the original-training requirement by themselves.

The 2027 edition reflects research reviewed on 22 October 2026. Five developers appear in an ordered Tier I, followed by ten established developers and fifteen specialists displayed alphabetically within their bands. The assessment is an editorial interpretation of documented technical programs. It is not a claim that all thirty organizations participated in a common laboratory test, nor a prediction that their relative model performance will remain fixed during 2027.

Market Overview

The largest general-purpose programs connect training with extensive distribution. OpenAI and Anthropic expose their models through managed developer interfaces, while Google DeepMind combines the Gemini program with the distinct Gemma release route. These arrangements give developers access to maintained systems but establish different boundaries around model inspection and operation. Managed access can simplify deployment while leaving architecture, data and weight availability partly or wholly under the developer’s control.

Open-weight development creates a different implementation choice. Meta, Alibaba, DeepSeek and other developers release artifacts that organizations can operate or adapt themselves. It also transfers substantial work to the deployer: a downloadable model still needs a supported serving environment, a clear license review and evaluation after any adaptation. Weight availability is therefore an important operating characteristic rather than a guarantee of complete scientific reproducibility.

Enterprise-oriented programs emphasize the relationship between models and controlled deployment. IBM’s Granite, Cohere’s Command and Microsoft’s Phi research illustrate different approaches to that problem. Transparency, retrieval fitness, compact-model design and deployment boundaries can matter as much as the highest result on a broad public leaderboard. A buyer needs to understand which part of a supplier’s proposition comes from the model itself and which part comes from surrounding software, services or integration work.

Visual and audio developers contribute independent foundation-model research. Black Forest Labs, Stability AI, Runway, Luma and ElevenLabs address generation tasks with different failure patterns and evaluation needs. Temporal inconsistency in video, unlicensed voice imitation and an incorrect factual answer are distinct technical and governance problems. Including those developers broadens the comparison only within the approved foundation-model scope; it does not turn unlike benchmark values into a synthetic universal intelligence score.

Corporate identity requires continuing attention. The xAI research operation became part of SpaceX during 2026, while Cohere and Aleph Alpha announced a business-combination agreement whose completion must be distinguished from its signing. The ranking follows the identifiable current developer and avoids awarding separate places to domestic and international brands representing the same organization.

Industry Trend — 2027

The direction toward 2027 is shaped by release choices as much as by larger training runs. Model programs are becoming easier to integrate in some respects and harder to compare in others: different reasoning settings, tool environments, modalities and serving configurations can change what a published result actually means. The table identifies developments visible in the reviewed material and the questions they create for model users.

2027 development	Documented direction	Technical implication
Model families rather than single releases	Developers maintain variants for different resource envelopes, input types and application tasks.	Select and record the actual artifact or endpoint; a family name is insufficient for reproducible evaluation.
Original training and adaptation coexist	Some developers publish both independently pretrained models and derivatives of external families.	Check provenance per family before attributing original-training capability or applying upstream terms.
Reasoning and tool use converge	Language models increasingly participate in applications that call external systems and execute multistep work.	Evaluate action authorization, failure recovery and task completion alongside answer quality.
Open-weight deployment remains differentiated	Weight access, training-data disclosure and unrestricted licensing remain separate properties.	Inspect each release’s license, model card and support assumptions before choosing self-managed operation.
Multimodal evaluation becomes more specialized	Speech, image and video programs expose different generation controls and evaluation methods.	Compare evidence within the relevant task and avoid treating visual preference as a language-reasoning result.
Transparency becomes an operational input	Training reports, system cards and model histories inform procurement and technical review.	Use disclosed limitations to design application tests; disclosure itself does not eliminate model risk.

These developments favor disciplined model inventories. An organization should be able to identify the developer, release, access route and evaluation configuration attached to every production use. Maintaining that record makes it possible to interpret results, investigate failures and decide whether a new release belongs in the same workflow.

Independent evaluation remains valuable when its task and conditions are understood. Leaderboards provide a useful view of particular capabilities, while transparency studies examine a different dimension of developer practice. Neither should be treated as a substitute for the other. This ranking uses those distinctions to interpret documentary evidence, preserving the difference between an observed benchmark result and the editorial assessment of a developer’s overall program.

Methodology — Core Eligibility Criteria

The ranking uses an evidence-based editorial assessment tailored to this technical market. Candidates were required to satisfy the following core conditions:

• An identifiable organization responsible for original pretraining of at least one reusable general-purpose foundation-model family.
• Evidence of continuing model maintenance, supported access or an active research and release program at the research cutoff.
• Language, multimodal, image, video and audio foundations may qualify when their capabilities extend beyond a single narrow sector application.
• A technical report, model card, release artifact or similarly substantive first-party account establishing development provenance.
• A clear distinction between original training, continued pretraining, fine-tuning, distillation and application-layer integration.
• One normalized developer entry per operating group; branded variants and absorbed businesses do not create duplicate places.
• Sufficient public information to identify the current operator, headquarters or accurately labeled operating location, and founding or institutional origin.
• Wrappers, pure resellers, inactive programs and unreleased prototypes without a qualifying maintained model are excluded.

Methodology — Ranking Factors

• Technical capability and model-program fitness — 30%: Original model development, modality-appropriate capabilities and depth within the maintained family, without pooling incompatible benchmark scales.
• Evaluation evidence and operational use — 25%: Inspectability of technical evaluations, version and task clarity, reproducible artifacts where available and evidence that the program is operational.
• Safety, reliability and data governance — 20%: Documented risk assessments, use boundaries, release controls, model limitations and data-handling practices, distinguished from guarantees of safe outputs.
• Developer usability and deployment access — 15%: Practical API or artifact access, deployment options, maintained interfaces and the relationship between model research and usable implementation.
• Technical transparency and responsible operation — 10%: Training provenance, architecture and license disclosure, clear release identity, accountable operations and candid treatment of limitations.

The five criteria carry weights of 30%, 25%, 20%, 15% and 10%, respectively. The private research record documents criterion-level editorial assessments and supporting sources. These are judgments about the documented offering, not scores from laboratory tests performed by The Economy Rankings. Missing material evidence is pursued before selection rather than silently treated as average performance. Licensing, sponsorship and likelihood of purchasing recognition do not influence eligibility, scores or tier placement.

Company Profiles and Further Reference

Names appearing in this ranking are linked to their corresponding organizational profiles in The Economy Wiki, where available. These profiles provide further background on principal activities, technology capabilities, product and service portfolios, market positioning and related rankings and analysis across The Economy Network.

Where an entry identifies a product or service line, its provider or operator is identified so that the scope of the organizational profile is clear. The Economy Wiki profiles are maintained as editorial reference pages and may be updated as new public information becomes available.

---

Tier I — Leading Foundation Model Developers

1. Google DeepMind

• Provider / operator: Google DeepMind / Google LLC
• Headquarters: London, United Kingdom (Google DeepMind)
• Provider founded / institutional origin: 2010 (DeepMind); Google DeepMind formed in 2023

Gemini and Gemma represent original multimodal and language-model research with both hosted and downloadable delivery routes. Multimodal pretraining addresses information distributed across different input forms, while the Gemma program provides a separate route for organizations seeking deployable model artifacts. Those are complementary development approaches with different operational consequences.

The combination of technical reports, developer APIs and cloud integration makes the family relevant to teams spanning interactive assistants and controlled model deployment. A buyer should compare the supported release, serving route and permission boundary that will actually enter production, rather than treating Google's entire research portfolio as one interchangeable model. The distinction between model research and distribution matters when several Google services expose related capabilities through different contracts and administrative controls.

Gemini and Gemma have different licenses and operating assumptions; Google's wider infrastructure is not itself a model-quality result. Tier I recognizes Google DeepMind's documented contribution: gemini and Gemma demonstrate original multimodal and language research across managed and deployable families.

2. Anthropic

• Provider / operator: Anthropic PBC
• Headquarters: San Francisco, California, United States
• Provider founded / institutional origin: 2021

Claude models support language reasoning, coding and tool-mediated work, with release-specific system cards and constitutional training research. The relationship between model behavior and a published constitution gives researchers a specific account of intended conduct. System cards also expose discrepancies between that intention and behavior found during testing.

Developers can combine messages, tool definitions and managed model access while examining disclosed safety evaluations and deployment decisions. For enterprises connecting Claude to internal systems, the meaningful unit of assurance includes the application's tool permissions and review gates as well as the underlying model. A published behavioral framework gives application owners a concrete basis for comparing intended conduct with failures found in their own evaluations.

System cards describe assessed versions and conditions; they cannot guarantee safe behavior across every tool permission or production prompt. Tier I recognizes Anthropic's documented contribution: claude combines general language, reasoning and coding within a maintained original model program.

3. OpenAI

• Provider / operator: OpenAI Group PBC / OpenAI Foundation
• Headquarters: San Francisco, California, United States
• Provider founded / institutional origin: 2015

GPT language and multimodal models combine instruction following, tool use and reasoning with managed developer access. A reusable model API lets an application supply its own context, request structured output and invoke explicitly defined tools. Those components make the model a development substrate rather than a finished answer engine.

The research proposition connects original foundation-model training with APIs used to build assistants, structured workflows and software tools; system cards document deployment evaluations. Buyers should evaluate answer quality together with tool authorization, fallback behavior and changes between model snapshots; those decisions remain outside a generic benchmark score. Tool schemas describe available actions; the surrounding application decides which systems a model can reach and when human review is necessary.

Closed weights limit external reproduction of training, and an API capability does not establish reliability in a buyer's own task. Tier I recognizes OpenAI's documented contribution: original training and broad multimodal/tool capability are documented, rather than inferred from ChatGPT distribution.

4. Alibaba

• Provider / operator: Alibaba Group / Qwen team
• Headquarters: Hangzhou, China
• Provider founded / institutional origin: 1999 (Alibaba Group)

Qwen develops original language and multimodal models with technical reports, public model cards and a broad range of deployment sizes. A family covering several model sizes can support different resource constraints without requiring every application to use the largest checkpoint. Multilingual and multimodal variants also let builders select a model whose inputs match their workflow.

Downloadable checkpoints and cloud access support multilingual applications, coding workflows and organizations needing choices about where inference runs. That flexibility increases the importance of recording the precise artifact, tokenizer and inference settings used in an evaluation, particularly when several Qwen variants share a familiar brand. Separate artifacts can suit different operational constraints, but a shared brand does not guarantee equivalent multilingual coverage or the same response behavior.

Model-family breadth creates version-selection work; a result from one Qwen release should not be transferred to another size or modality. Tier I recognizes Alibaba's documented contribution: qwen provides substantive original training across language, multimodal and different resource envelopes.

5. Meta

• Provider / operator: Meta Platforms, Inc.
• Headquarters: Menlo Park, California, United States
• Provider founded / institutional origin: 2004

Llama provides pretrained and instruction-adapted language models, with downloadable weights and documented model architecture and use policies. Downloadable parameters change the implementation decision: teams can choose infrastructure, adaptation methods and runtime controls. Model cards provide a point of reference for the training assumptions inherited by those downstream deployments.

Teams can inspect model cards, adapt supported releases and operate inference in their own environments instead of depending exclusively on a hosted endpoint. The practical obligation is to preserve provenance through adaptation and verify that license obligations, application safeguards and evaluation data remain appropriate after the model has been modified. Self-hosting also makes version retention possible, letting engineering teams compare a proposed upgrade against an unchanged model in their existing environment.

Open weights do not disclose the complete training corpus, and each release's license and safety requirements still apply. Tier I recognizes Meta's documented contribution: llama establishes substantial original pretraining and downstream general-purpose deployment capability.

---

Tier II — Established Foundation Model Developers

(Alphabetical order)

Amazon

• Provider / operator: Amazon.com, Inc. / Amazon AGI
• Headquarters: Seattle, Washington; Arlington, Virginia, United States
• Provider founded / institutional origin: 1994

Nova is Amazon's own family of foundation models, with language and multimodal capabilities and model-specific service-card documentation. Service cards identify intended uses, limitations and responsible deployment considerations around a managed model family. They give procurement and engineering teams a common starting point for discussing the boundary between model behavior and service operation.

Managed access through Amazon's AI services gives enterprise builders a route to combine original models with application controls, evaluation and deployment operations. Organizations already using AWS can examine integration convenience separately from model selection, testing whether Nova's actual outputs satisfy the business task before standardizing on a deployment route. Managed model delivery places service administration alongside application engineering, making account controls and release selection part of the model user's operating process.

The evaluation separates Nova development from Bedrock's resale of other developers' models and distinguishes current releases from announced roadmaps. Tier II recognizes Amazon's documented contribution: nova is an original multimodal foundation-model program distinct from the Bedrock marketplace.

Black Forest Labs

• Provider / operator: Black Forest Labs, Inc. / Black Forest Labs GmbH
• Headquarters: Freiburg, Germany; San Francisco, United States (research locations)
• Provider founded / institutional origin: 2024

FLUX is an original visual foundation-model family developed by researchers with a documented history in latent diffusion and flow-based generation. Flow-based visual generation and selected downloadable artifacts give FLUX a substantive model-development identity. The distinction between managed and downloadable variants also makes release engineering and access terms part of the technical comparison.

The lab combines API access with selected downloadable releases, allowing creative developers to choose between managed inference and greater deployment control. Creative teams should evaluate prompt adherence and edit behavior using their own production material, preserving a record of the specific variant rather than treating every image produced under the FLUX name as equivalent. A creative model's development value includes how readily an application can control and repeat a workflow, alongside the visual qualities of selected outputs.

Different FLUX variants carry different access and licensing terms, while visual preference tests do not establish general language or coding ability. Tier II recognizes Black Forest Labs's documented contribution: fLUX establishes original flow-based visual foundation-model research and controllable generation.

Cohere

• Provider / operator: Cohere Inc.
• Headquarters: Toronto, Canada; announced Berlin co-headquarters subject to transaction completion
• Provider founded / institutional origin: 2019

Command models focus on enterprise language tasks, retrieval-grounded responses and multilingual applications, with model and deployment documentation. Retrieval-oriented model development addresses a practical enterprise requirement: using institutional information within an answer while preserving a clear deployment boundary. That focus creates a distinct position alongside labs primarily oriented toward consumer assistants or broad creative generation.

The developer combines original model training with enterprise deployment choices, making it relevant to organizations examining data control alongside application capability. A successful implementation still requires a maintained source collection, appropriate retrieval configuration and authorization checks; a model trained for grounded responses cannot repair an inaccessible or poorly governed knowledge base. Retrieval-oriented development is particularly relevant when an organization's useful knowledge changes more quickly than a pretrained model's general information about the world.

The announced Aleph Alpha combination is not represented as closed without confirmation, and retrieval success still depends on the customer's corpus and permissions. Tier II recognizes Cohere's documented contribution: command documents original enterprise language-model development with retrieval-oriented task fitness.

DeepSeek

• Provider / operator: DeepSeek
• Headquarters: Hangzhou, China
• Provider founded / institutional origin: 2023

DeepSeek's original mixture-of-experts models and reasoning research disclose architecture and training methods through technical reports and released weights. Mixture-of-experts design separates the model's overall parameter capacity from the subset used for a particular token. The published architecture gives researchers something more concrete to inspect than a general efficiency slogan.

The family is relevant to developers studying efficient pretraining, reasoning-oriented post-training and self-managed deployment of general-purpose language systems. Downstream teams still need workload-specific serving estimates, because model architecture alone does not determine latency, memory use or total operating cost under their concurrency and context requirements. Inspecting the base model separately from reasoning adaptations helps distinguish architectural efficiency from additional inference work performed while producing an answer.

A reasoning adaptation must be distinguished from its pretrained base, and published evaluations require attention to harness and sampling settings. Tier II recognizes DeepSeek's documented contribution: the original MoE training program and disclosed architecture establish substantive model capability and research depth.

ElevenLabs

• Provider / operator: ElevenLabs Inc.
• Headquarters: London, United Kingdom (European headquarters and founding base)
• Provider founded / institutional origin: 2022

ElevenLabs develops original speech and audio models, beginning with in-house speech synthesis and extending across audio generation and recognition. Speech generation depends on pronunciation, prosody, language coverage and controllable delivery, not only whether a sentence is intelligible. ElevenLabs connects original audio-model research with interfaces that make those characteristics accessible to application developers.

Developer APIs and documented voice workflows make its model research relevant to accessibility, localization, conversational interfaces and professional audio production. For deployment, consent and identity protection are practical parts of model fitness, particularly when a workflow can reproduce recognizable voices or distribute generated speech to a broad audience. Audio applications need to evaluate what is said and how it is delivered, including pronunciation, pacing and the representation of a recognizable speaker.

Voice quality, speaker consent and misuse controls require modality-specific assessment; speech-model recognition is not a claim of frontier text-model reasoning. Tier II recognizes ElevenLabs's documented contribution: original in-house speech synthesis and subsequent audio research establish substantive modality-specific model depth.

IBM

• Provider / operator: International Business Machines Corporation
• Headquarters: Armonk, New York, United States
• Provider founded / institutional origin: 1911

Granite develops original enterprise-oriented language models and associated research, with openly available model artifacts and technical documentation. Open model artifacts are particularly useful when an organization wants to maintain its own validation process and deployment inventory. Granite's research material makes training and release practices part of the product's technical identity.

The proposition combines model training transparency with enterprise deployment tooling, supporting organizations that need to inspect licensing and govern model operation. For governed enterprise work, that inspectability helps connect model selection with approval procedures, but a transparent model still requires testing against the documents, languages and errors relevant to the organization. A well-documented release helps reviewers connect a model artifact with its permitted use, internal approval record and later changes in production behavior.

Enterprise positioning does not establish superiority on every general reasoning task; the released model and deployment configuration remain material. Tier II recognizes IBM's documented contribution: granite supplies original enterprise-oriented language training and maintained model artifacts.

Microsoft

• Provider / operator: Microsoft Corporation / Microsoft AI
• Headquarters: Redmond, Washington, United States
• Provider founded / institutional origin: 1975

Phi research establishes original small-model training, emphasizing curated data and compact language or multimodal model designs. Compact-model research makes the relationship between training-data design and model size particularly visible. A smaller artifact can be relevant where a task has constrained resource or deployment requirements, even when a larger general model remains stronger elsewhere.

Technical reports and downloadable models give application builders a way to examine smaller-model tradeoffs within a wider enterprise developer environment. The useful comparison is task fitness under those constraints, with the Phi release evaluated on its own documentation rather than borrowing capability claims from unrelated Microsoft-hosted models. A compact model is most useful when its supported task, hardware envelope and error tolerance form a coherent application design around the artifact.

The entry evaluates Microsoft's own model development, rather than treating access to third-party Azure models as original training. Tier II recognizes Microsoft's documented contribution: phi reports establish original compact-model training with an identifiable data-design approach.

Mistral AI

• Provider / operator: Mistral AI
• Headquarters: Paris, France
• Provider founded / institutional origin: 2023

Mistral develops original language and multimodal models, combining open-weight releases with managed models and enterprise deployment options. The coexistence of deployable models and managed services gives implementation teams more than one control boundary. An enterprise can examine where model operation should sit relative to its data, infrastructure and internal support responsibilities.

The product architecture supports teams choosing between direct APIs and greater control over model operation, while documentation separates model families and access routes. Model selection therefore requires reading the terms attached to the actual release and separating optional hosting services from the technical properties of the foundation model itself. An organization can value model portability while still choosing managed delivery for particular applications, provided its operating responsibilities are understood in advance.

Weight availability and commercial terms vary by model; the assessment does not assume every Mistral release has identical openness. Tier II recognizes Mistral AI's documented contribution: original language and multimodal training supports a technically substantial model portfolio.

Moonshot AI

• Provider / operator: Moonshot AI
• Headquarters: Beijing, China
• Provider founded / institutional origin: 2023

Kimi develops original large language models with mixture-of-experts pretraining research and reasoning or agent-oriented post-training. The Kimi research account includes choices about optimization and the utility of training tokens, making the training process an identifiable subject of evaluation. Released artifacts also support closer inspection of the resulting language-model behavior.

Technical reports and available model artifacts support examination of architecture, training design and deployment choices for long-form language applications. For applications handling long documents or tools, buyers should test information access and action selection across realistic tasks, including failures that an overall reasoning score might obscure. The relationship between training design and downstream behavior remains an evaluation question; architectural descriptions provide context rather than a substitute for testing representative tasks.

Context capacity does not itself prove accurate retrieval throughout a document, and benchmark claims must be tied to the released Kimi version. Tier II recognizes Moonshot AI's documented contribution: kimi combines original MoE pretraining with documented language and agent-oriented research.

NVIDIA

• Provider / operator: NVIDIA Corporation
• Headquarters: Santa Clara, California, United States
• Provider founded / institutional origin: 1993

Nemotron includes original pretrained families alongside adaptations of external architectures, with public recipes, datasets and model cards. A published pretraining recipe gives researchers visibility into the sequence that produced a base model, including how training infrastructure and data preparation connect. This is a stronger provenance signal than an instruction-tuned checkpoint without an identified base.

The original-training evidence includes Nemotron base models and the Nemotron 3 pretraining recipe, connecting model research with a substantial development stack. NVIDIA's adaptation work remains relevant engineering, but separating those lineages prevents the original-training criterion from becoming a reward for repackaging another developer's model under a new name. Publishing training machinery can help researchers investigate how a model was produced, even when exact reproduction requires resources beyond an individual development team.

Llama-derived Nemotron releases are not treated as evidence of independent base-model invention; provenance is reviewed separately for each family. Tier II recognizes NVIDIA's documented contribution: nemotron original pretraining is explicitly established separately from Llama-derived adaptation work.

---

Tier III — Specialist Foundation Model Developers

(Alphabetical order)

Apple

• Provider / operator: Apple Inc.
• Headquarters: Cupertino, California, United States
• Provider founded / institutional origin: 1976

Apple develops original on-device and server foundation models, publishing architecture, training and evaluation reports for its model program. On-device execution makes model size, memory use and the relationship with local data central design questions. Server models address a different capacity envelope, while the published reports explain how the two development paths relate.

Its technical focus connects compact models with device constraints, inference optimization and an operating environment designed around integrated user experiences. The resulting proposition matters to applications built within Apple's supported environment; it should be evaluated against those constraints rather than as an unrestricted replacement for every hosted foundation-model API. The resource envelope is an explicit part of the model proposition, connecting training decisions with the practical constraints of inference on supported devices.

Apple's developer access and supported tasks differ from unrestricted model hosting; platform integration should not be mistaken for universal model availability. Tier III recognizes Apple's documented contribution: original device and server models address a coherent resource-aware foundation-model program.

Baidu

• Provider / operator: Baidu, Inc.
• Headquarters: Beijing, China
• Provider founded / institutional origin: 2000

ERNIE is Baidu's original foundation-model program, rooted in language pretraining and extended to multimodal and reasoning applications. Knowledge-oriented pretraining and Chinese-language development give ERNIE an identifiable research lineage. Its model program can be examined separately from the search and advertising businesses that distribute related services to a large audience.

Technical publications and platform access make the family relevant to Chinese-language applications and enterprise systems requiring integration with Baidu's AI tooling. Enterprise developers should verify language coverage, supported interfaces and evaluation conditions for the particular ERNIE release, especially when an application combines retrieval, model reasoning and domain-specific information. The original pretraining lineage is evaluated independently from the applications and enterprise integrations that subsequently specialize the model for a particular user.

The assessment concerns model development and documented access, without inferring cross-language superiority from vendor-reported domestic evaluations. Tier III recognizes Baidu's documented contribution: eRNIE research establishes original language pretraining and multimodal development.

ByteDance

• Provider / operator: ByteDance Ltd. / Seed
• Headquarters: Beijing, China (core operating base)
• Provider founded / institutional origin: 2012

Seed develops original language, vision and generative-media foundation models, with a portfolio of technical reports and research publications. Seed's publication program covers model training as well as systems research, making the engineering behind model development visible beyond a consumer application's interface. This supports evaluating the organization as a developer rather than merely a distributor.

The program combines multimodal pretraining research with developer-facing delivery, giving application teams access to model capabilities beyond ByteDance's consumer apps. The relevant evidence remains release-specific: a strong visual demonstration or widely used application cannot substitute for a documented language evaluation, training account or enterprise deployment condition. A documented research program gives developers a way to distinguish newly trained capabilities from interface features added around a previously available model service.

Consumer distribution does not establish independent model quality, and public evidence varies between open technical reports and proprietary product releases. Tier III recognizes ByteDance's documented contribution: seed demonstrates original multimodal pretraining and generative development across a broad research portfolio.

Huawei

• Provider / operator: Huawei Technologies Co., Ltd.
• Headquarters: Shenzhen, China
• Provider founded / institutional origin: 1987

Pangu is Huawei's original foundation-model program, with language and multimodal training research connected to its AI computing environment. The Pangu architecture separates general foundation models from industry and scenario adaptations. That hierarchy is useful for identifying the reusable training contribution without counting every downstream specialization as a separate general-purpose model developer.

The model family serves general AI development and enterprise adaptation, while published work distinguishes model research from the underlying accelerator and cloud business. Enterprise users should distinguish base-model capability from the value added by industry data and integration work, then examine whether the relevant deployment route meets their operational and governance requirements. The distinction between general and adapted layers helps enterprise teams locate the source of a capability before assigning responsibility for a deployment or update.

Industry adaptations are not counted as separate developers, and hardware availability does not itself establish a model's independent task performance. Tier III recognizes Huawei's documented contribution: pangu distinguishes general foundation training from industry and scenario adaptation.

Luma AI

• Provider / operator: Luma AI, Inc.
• Headquarters: Palo Alto, California, United States
• Provider founded / institutional origin: 2021

Luma develops original visual and video foundation models, with Ray models delivered through its creative environment and developer services. A production-oriented video model needs to connect generation controls with the formats and iteration habits of creative work. Luma's Ray program makes that relationship visible through its model releases and the surrounding creative environment.

The research proposition connects video generation with visual reasoning and production-oriented output controls, giving creative teams an alternative foundation-model source. The relevant evaluation includes temporal coherence and controllability within the supported workflow, while broader claims about simulated worlds remain hypotheses requiring more specific evidence than attractive generated footage. Creative iteration depends on predictable controls and usable outputs, which makes the surrounding generation workflow relevant without turning interface convenience into model superiority.

Promotional descriptions of reasoning and world modeling are treated as developer claims; the ranking makes no independent physics-understanding certification. Tier III recognizes Luma AI's documented contribution: ray establishes original visual model development with production-oriented generation controls.

MiniMax

• Provider / operator: MiniMax Group Inc.
• Headquarters: Shanghai, China
• Provider founded / institutional origin: 2022 (official company account; predecessor operating entity formed in 2021)

MiniMax trains original language, speech and video models, combining mixture-of-experts language research with multimodal generation. Language, speech and video place different demands on training data and evaluation. MiniMax's program is therefore assessed as a developer portfolio with modality-specific evidence, rather than by adding unrelated published benchmark values together.

Its public model program and developer APIs support applications spanning coding agents, voice experiences and creative production, giving the lab a cross-modality identity. Its access routes are relevant to teams assembling applications across media, but each model's terms, inference interface and safety restrictions need to be checked independently before those components are combined. Cross-modality breadth can matter to application builders, provided the assessment keeps each model's inputs, outputs and intended operating conditions separate throughout technical review.

The official founding account is distinguished from predecessor incorporation, and the article avoids combining unrelated modality benchmarks into one capability claim. Tier III recognizes MiniMax's documented contribution: original language, audio and visual families demonstrate a substantive cross-modality training program.

Reka AI

• Provider / operator: Reka AI, Inc.
• Headquarters: San Francisco Bay Area, California, United States
• Provider founded / institutional origin: 2022

Reka develops original multimodal foundation models, with Core research covering language and visual or audio understanding. Training multimodal models from scratch establishes a provenance boundary that can be examined through the developer's technical report. The work addresses reusable understanding across input types rather than only building a user interface around another lab's endpoint.

The program combines research publications with model services, offering application builders a developer whose primary identity includes training general-purpose multimodal systems. For organizations evaluating a smaller model supplier, the next question is whether its documented access, support arrangements and task coverage match the intended application's requirements throughout a continuing deployment. A clearly stated original-training account lets users examine the developer's contribution before considering distribution arrangements or the capabilities added by a downstream application.

A smaller public deployment record limits broad operational conclusions, while reported multimodal evaluations remain specific to their test conditions. Tier III recognizes Reka AI's documented contribution: the Core report explicitly establishes original multimodal training from scratch.

Runway

• Provider / operator: Runway AI, Inc.
• Headquarters: New York, New York, United States
• Provider founded / institutional origin: 2018

Runway trains original video-generation models, building on published generative research and successive Gen-family releases. Video models must organize change across frames as well as the content of an individual image. Runway's research and product progression therefore address a different technical problem from a text assistant or a still-image generator.

Its model program connects temporal visual generation with practical production interfaces and developer access, supporting teams creating or transforming moving images. Production users should inspect motion consistency, reference control and failure behavior in the kind of scene they intend to create, rather than extrapolating from a carefully selected promotional sequence. Temporal generation introduces failure modes that a still-image evaluation cannot capture, making sequence-level inspection central to a useful technical review of the model.

Creative demonstrations are not controlled evidence of universal physical understanding, and video-generation quality cannot be directly pooled with text-reasoning results. Tier III recognizes Runway's documented contribution: the Gen research program addresses original temporal visual generation and reference-conditioned creation.

SenseTime

• Provider / operator: SenseTime Group Inc.
• Headquarters: Hong Kong SAR, China
• Provider founded / institutional origin: 2014

SenseNova brings together original foundation-model development in language and multimodal generation with SenseTime's AI research infrastructure. SenseNova's foundation-model identity sits alongside a much longer history of specialized AI applications. Separating the two makes it possible to consider reusable model capabilities without turning established customer relationships into automatic evidence of general-purpose reasoning quality.

The program offers general-purpose model capabilities to developers and enterprises, distinct from the company's narrower computer-vision applications and industry solutions. The buyer's evaluation should identify the supported model service and adaptation pathway, with institutional infrastructure providing operational context rather than acting as a substitute for task-level evidence. Reusable models and industry implementations play different roles, so the profile keeps general training capability separate from the delivery of a specialized enterprise solution.

Infrastructure scale and historical vision expertise do not automatically establish language-model leadership; the comparison stays with documented model capabilities. Tier III recognizes SenseTime's documented contribution: senseNova supplies reusable foundation-model development distinct from narrower vision applications.

Stability AI

• Provider / operator: Stability AI Ltd.
• Headquarters: London, United Kingdom
• Provider founded / institutional origin: 2019

Stable Diffusion and associated creative-model research establish original generative training across image and other media modalities. Image-generation developers need to assess prompt interpretation, editing control, output characteristics and deployment requirements as a connected workflow. Stable Diffusion's available artifacts make local operation and adaptation practical subjects for technical examination.

Downloadable model releases and deployment documentation give developers substantial flexibility to build image-generation workflows and inspect model-specific terms. The scope remains general visual generation: the model's usefulness as a creative foundation is meaningful without implying that the same artifact can perform the language tasks used to evaluate conversational systems. The availability of an artifact makes adaptation possible, but the adapted result needs a fresh review of output behavior and applicable release conditions.

An image model is assessed for visual generation and release practice, not assigned an invented score on a language-reasoning benchmark. Tier III recognizes Stability AI's documented contribution: original visual generative training establishes foundation-model depth in an allowed modality.

StepFun

• Provider / operator: Shanghai Jieyue Xingchen Intelligent Technology Co., Ltd.
• Headquarters: Shanghai, China
• Provider founded / institutional origin: 2023

StepFun develops original language and multimodal model families, alongside separately identified adaptations and vision-language integrations. The distinction between original pretraining and multimodal adaptation is especially important when a developer publishes several kinds of model. StepFun's broader program includes original families, while individual integrations can draw on externally developed components.

Published research spans language, image, video and audio, giving the developer a substantive model-training identity beyond a single assistant interface. Tracking those lineages helps users understand what is attributable to the developer, which upstream licenses apply and which results concern the actual model being deployed. An upstream component can contribute useful capability without becoming an original invention of the organization that later integrates it into a larger model system.

A vision-language model using an external decoder is not taken alone as original base-model proof; the broader Step pretraining program establishes eligibility. Tier III recognizes StepFun's documented contribution: step original model families are distinguished from vision-language integrations using external components.

Technology Innovation Institute

• Provider / operator: Technology Innovation Institute, Advanced Technology Research Council
• Headquarters: Abu Dhabi, United Arab Emirates
• Provider founded / institutional origin: 2020

Falcon includes original pretrained language-model families, including hybrid designs that combine attention and state-space modeling. Hybrid language-model designs explore how state-space processing and attention can share work within an architecture. Public releases allow researchers to study the resulting deployment tradeoffs rather than relying only on a managed interface.

Public model releases and documentation make the institute relevant to developers studying alternative architectures and deploying general-purpose language models outside a single hosted service. The institute's inclusion recognizes a reusable model-development contribution; buyers should still examine available implementation support and license obligations before using a research release as the basis of an operational service. A public research release can widen implementation choices, while the availability of community artifacts remains distinct from a contractual promise of ongoing operational support.

The institute's institutional research setting gives Falcon a distinct role among the commercial developers in this comparison. Tier III reflects accessible model artifacts and an inspectable research direction, with operational support assessed separately from the scientific contribution.

Tencent

• Provider / operator: Tencent Holdings Limited / Hunyuan team
• Headquarters: Shenzhen, China
• Provider founded / institutional origin: 1998

Hunyuan spans original language and generative-media model research, supported by model releases and technical disclosures. A broad multimodal program can reuse research infrastructure while producing models for substantially different tasks. The existence of language and media releases is evidence of development breadth, while their results still need separate interpretation.

The program links model development with Tencent Cloud access and selected downloadable artifacts, serving text, image, video and related application needs. A production team should begin with the artifact or endpoint suited to its inputs and follow its own permissions, acceptable-use and integration requirements through to the finished application. Keeping the model artifact and modality visible prevents the breadth of a corporate AI portfolio from obscuring the requirements of one production task.

Capabilities and licenses differ across Hunyuan modalities, so image-generation progress is not treated as a language-reasoning benchmark result. Tier III recognizes Tencent's documented contribution: hunyuan develops original language and generative-media models rather than only adapting an assistant interface.

xAI / SpaceXAI

• Provider / operator: SpaceXAI LLC, a SpaceX subsidiary
• Headquarters: Palo Alto, California, United States (operating address)
• Provider founded / institutional origin: 2023 (xAI origin)

Grok is an original general-purpose model family with developer access and release-related capability and safety information. A change in corporate ownership does not erase the model program's training lineage, but it does change the identity that buyers must track in contracts and operating records. Grok remains the relevant family for this comparison.

The evaluated research operation is the former xAI business within SpaceX, rather than SpaceX's launch activity or the visibility of the X distribution channel. The entry therefore follows the active AI developer and its documented model access, while treating planned compute expansion and distribution through social media as separate considerations from assessed technical capability. Following the operating developer through an acquisition keeps model provenance intact while making the current organizational and contracting boundary clear to prospective users.

Ownership and branding changed during 2026; model capability, safeguards and availability are assessed separately from the parent company's infrastructure ambitions. Tier III recognizes xAI / SpaceXAI's documented contribution: grok is a maintained original general-purpose family, evaluated apart from distribution through X.

Z.ai / Zhipu AI

• Provider / operator: Z.AI Co., Ltd. / Beijing Zhipu Huazhang
• Headquarters: Beijing, China
• Provider founded / institutional origin: 2019

GLM represents an original language-model development program with published pretraining research and maintained model releases. The GLM lineage connects an explicit pretraining method with later language and multimodal releases. Publishing model artifacts gives outside developers a route to examine behavior and adapt supported models within their own environments.

The family connects bilingual model research with coding, multimodal and tool-use applications, supported by public artifacts and developer access. Using both the international and domestic brand in one entry avoids counting the same development organization twice, while preserving a clear link between the named developer and its current model family. Normalizing the developer's identity across brands makes the model history easier to follow and preserves a consistent basis for comparing successive research releases.

Current branding is normalized to one developer entry; earlier GLM papers establish provenance but do not alone validate the capabilities of later releases. Tier III recognizes Z.ai / Zhipu AI's documented contribution: the GLM lineage establishes original pretraining with maintained language and multimodal development.

---

Remarks

Tier classification reflects the strength of documented developer programs under the five stated factors. It is broader than the position of one model on one leaderboard, and it should be read with the modality and deployment scope of each profile. An organization selecting a speech engine, a compact device model or a visual-generation base can reasonably prefer a specialist whose program matches its particular task.

The ranking does not certify that a developer’s models will be accurate, safe or economical in every application. Production performance depends on the chosen release, inference configuration, retrieved information, tool environment and user behavior. Where a developer reports an evaluation, its conditions remain part of the claim. No numerical benchmark values, savings or productivity improvements are invented or inferred from the editorial tiers.

Headquarters and founding fields identify the provider or institution described in each entry, with explicit labels where a laboratory, predecessor or operating address differs from a conventional corporate headquarters. Announced transactions are distinguished from confirmed acquisitions. Organizational profiles linked through The Economy Wiki are planned editorial reference destinations and may evolve as ownership, branding and model access change.

The useful follow-on work is a task-specific evaluation with documented acceptance criteria. Buyers should test representative inputs, retain a record of the model version and examine failure handling before expanding access. A strong research program can provide a promising starting point, but the final deployment decision depends on the organization’s information, risk tolerance and ability to maintain the application over time.

---

Recognition

Inclusion in the Top 30 Foundation Model Developers 2027 ranking is an editorial determination of The Economy Rankings and does not require licensing, advertising, sponsorship, or other commercial participation.

Ranked organizations may factually refer to their inclusion in the ranking in their own communications. When describing the result, entities should accurately reflect the tier structure and methodology used in the published ranking.

How the ranking should be interpreted

• Tier I represents the Top 5 entities, and the published order within Tier I reflects the ranking order.
• Tier II represents entities ranked within the Top 15, following Tier I. Entities within Tier II are displayed alphabetically; their displayed order should therefore not be interpreted as an individual numerical ranking.
• Tier III represents entities ranked within the Top 30, following Tiers I and II. Entities within Tier III are also displayed alphabetically, and their displayed order should not be interpreted as an individual numerical ranking.
• An entity's tier, rather than its alphabetical position within Tier II or Tier III, should therefore be used when describing its standing.

Referencing the ranking

Depending on the entity's published tier, appropriate factual descriptions may include:

• Tier I: “Ranked Tier I” or “Ranked among the Top 5”
• Tier II: “Ranked Tier II” or “Ranked among the Top 15”
• Tier III: “Ranked Tier III” or “Ranked among the Top 30”

Entities should not describe an alphabetical position within Tier II or Tier III as a specific numerical rank.

Use of The Economy Rankings recognition materials

Editorial inclusion in a ranking does NOT by itself grant permission to use The Economy Rankings badges, seals, logos, official recognition graphics, licensed quotations, or other proprietary recognition materials.

Organizations wishing to use official The Economy Rankings recognition materials in corporate websites, marketing materials, investor communications, client presentations, social media, press releases, or other external communications should refer to the applicable licensing terms and usage policies:

• The Economy Rankings License Structure — available licence levels and permitted recognition uses
  The Economy Rankings License Structure | The Economy Rankings
• The Economy Rankings License Policy — licensing conditions governing use of The Economy Rankings recognition, marks, and related materials
  The Economy Rankings License Policy | The Economy Rankings
• Recognition Use Guide — guidance on permitted wording, presentation, attribution, and use of ranking recognition
  Recognition Use Guide | The Economy Rankings

Purchasing or holding a recognition-materials licence does not confer inclusion or a higher tier.

Recognized organizations may reference the designation in:

• corporate websites
• investor communications
• marketing materials
• client presentations

Licensing inquiries:
[email protected]