8 segments · 134 companies · as of Sep 18, 2026
Frontier Model Labs
The model builders, regional challengers and listed investors funding them
Where it sits
Fed by: Cloud & AI CapitalFeeds: Applications & Agents · Physical AI
Foundation-model labs turn compute, data and research into reusable intelligence. Their economics depend on training costs, inference demand and distribution. Public exposure comes through in-house labs, newly listed developers, minority stakes and funds holding otherwise private companies.
Why now
New model generations and large 2026 financings reset the field. SpaceX now owns xAI and Cursor, Chinese labs trade in Hong Kong, and Korea has narrowed its sovereign-model contest. Financing commitments and pending acquisitions require careful separation from completed ownership changes.
US Frontier & In-House Labs
Independent US labs compete with technology groups that train their own models. Reasoning, coding and distribution through cloud APIs determine adoption; early research labs are distinguished from companies with released, generally available flagships.
GPT-6 Astra and GPT-5.6 Sol power ChatGPT and Codex; Azure, AWS and NVIDIA expand compute access
leaderClaude Fable 5.1 serves AWS, Azure and Google Cloud; Mythos 5.1 remains restricted to trusted access
leaderDeepMind trains Gemini 3.8 Flash for Google Cloud and Antigravity; 3.8 Flash Cyber has controlled access
majorSuperintelligence Labs develops Muse Spark 1.3 for Meta AI and app distribution
majorxAI unit trains Grok 4.5 on Colossus; Grok APIs and the acquired Cursor business extend distribution
majorMAI-Thinking-1 preview and MAI-Code-1.1 complement partner models across Microsoft AI products
challengerAGI lab develops Nova 2 Lite and Sonic for Bedrock; Nova 2 Pro and Omni remain preview offerings
majorAFM 3 combines on-device and cloud models with Google technology for Apple Intelligence and Siri
majorNemotron 3.5 Lightning supplies open enterprise reasoning models through NVIDIA software and partner clouds
majorGranite 4.2 offers Apache-licensed reasoning models through watsonx, Hugging Face and hybrid deployments
nicheDevelops superintelligence systems with NVIDIA compute support; no publicly released model verified by cutoff
emergingTinker provides fine-tuning and training APIs for researchers; a proprietary general-purpose flagship remains unverified
emergingDevelops open frontier models using a reported SpaceX training-compute agreement; public flagship remains unverified
emergingPersimmon v0.1 research preview targets human collaboration; its September release is not general availability
emergingBuilds scientific models and automated experiments with a16z backing; no public flagship verified
emergingINTELLECT-3 demonstrates open large-scale RL training; shared environments and compute support model developers
emergingMercury diffusion language models generate text in parallel for low-latency coding assistants and APIs
emergingTrinity open-weight models and model adaptation target enterprise deployments on customer-controlled infrastructure
emergingHermes open models and distributed post-training research supply developer-controlled reasoning and tool use
emergingCogito v2.1 reasoning models use iterative distillation and publish open weights for developer deployment
emergingPythia model suites and open training research support reproducible language-model science and evaluation
emergingGameplay-trained world models learn spatial reasoning and action prediction for interactive environments
emerging
Sources: persimmon.humansand.ai, openai.com, anthropic.com, blog.google, cursor.com, ibm.com
China Foundation Labs
Chinese model developers span listed pure plays, internet-platform labs and private challengers. Open-weight releases, domestic cloud distribution and efficient inference matter alongside export constraints; announced IPO plans alone do not change listing status.
V4.1-Flash supplies multimodal reasoning APIs; retired V4-Pro endpoints route to Flash from September 14
leaderQwen3.8 Max and Omni Flash feed Model Studio, coding agents and Alibaba Cloud distribution
leaderGLM-5.3 and 5.3-Flash power Z.ai coding plans and enterprise agents after the January HK listing
majorM3 supplies coding and agent reasoning; H3 adds open-weight video after the January HK listing
majorKimi K3 offers 1M-context coding and visual reasoning APIs; Alibaba and Tencent are financial backers
majorSeed 2.0 powers Doubao and Volcano Engine; Seedance 2.5 extends the family into video creation
majorHunyuan Hy4-preview opens large-model weights; Tencent Cloud connects the family to enterprise developers
majorERNIE 5.1 supplies reasoning and agents through Qianfan and the ERNIE developer platform
majorStep-3.7-Flash serves coding and tool-using agents through its APIs, OpenRouter and NVIDIA NIM
challengerLongCat-2.0 offers trillion-parameter agent and coding APIs alongside Meituan internal deployment
challengerMiMo-V2.5-Pro opens 1T-class coding and agent weights through MiMo APIs and AI Studio
challengerInclusionAI releases Ling-3.0-Flash and Tiny open reasoning models; Alibaba is a minority owner
challengerSenseNova U1.5 Lite supplies open multimodal image reasoning and editing through SenseCore
challengerSpark X1.5 supports speech and enterprise AI; Smart Blackboard connects it to Chinese education customers
nicheSkywork agents and SkyReels video models serve research, document creation and controllable media generation
nicheBaichuan-M3 and M3-Plus APIs target medical reasoning, consultation and retrieval from clinical sources
nichePangu 5.5 and openPangu models run on Ascend and Huawei Cloud for Chinese industrial and sovereign customers
majorYi open models and Wanzhi enterprise adaptation commercialize Chinese-language reasoning and knowledge work
nicheEmu multimodal and other open model families supply reusable Chinese research weights and training methods
nicheInternLM and InternVL publish open language and vision-language weights for researchers and application builders
nicheInfini-Megrez language and multimodal models target efficient deployment on heterogeneous Chinese accelerators
niche
Sources: platform.baichuan-ai.com, static.stepfun.com, docs.z.ai, minimax.io, platform.moonshot.ai, deepseek.com
Korean Foundation Labs
Korean labs compete on local language, enterprise deployment and national model development. The August 2026 sovereign-program evaluation advances SK Telecom, Upstage and LG; other model developers remain commercially relevant despite leaving or not joining that contest.
LG AI Research trains K-EXAONE 2.0; advances in the August sovereign-model evaluation
leaderSolar Open 2 and Pro 4 serve enterprise agents and OpenRouter; advances in the August sovereign round
leaderA.X K2 expands Korean reasoning; advances in the August sovereign round and maintains Anthropic ties
majorHyperCLOVA X SEED Think/Omni serve Naver Cloud; the sovereign-program team exits in January
majorMi:dm K 2.5 Pro targets Korean B2B agents; develops commercially outside the three August finalists
majorKanana-2 opens six-language agent models; operates outside the August sovereign finalist group
majorNC AI develops VAETKI and VARCO Vision for industry; its sovereign-program team exits in January
challengerMoreh-backed Motif 3 opens MIT-licensed weights; the team is eliminated in the August sovereign round
challengerKonan LLM ENT-11 combines reasoning with its own search engine for Korean enterprise agents
nicheLUXIA3 LLM, Deep and VLM serve Korean enterprises through LUXIA Cloud and on-premises LUXIA On
nicheTri-70B-preview-SFT publishes Korean, English and Japanese model weights for open-model research
emergingGauss2 language, code and image models support internal engineering and enterprise device workflows
majorMathGPT, co-developed with Upstage and KT, specializes in mathematical reasoning for QANDA learning products
niche
Sources: newspim.com, lgresearch.ai, news.sktelecom.com, upstage.ai, navercorp.com, konantech.com
Japan & Taiwan Labs
Local-language models and industrial research connect Japanese and Taiwanese groups to foundation-model development. Private labs work with banks and manufacturers; listed telecom and IT parents provide enterprise distribution, with model adaptation distinguished from original pretraining.
Namazu v1.0 adapts Kimi K2.6 for Japanese business; MUFG partnership connects research to banking
leaderPLaMo 2.2 Prime serves Japanese enterprises; Snowflake distribution complements Toyota-linked research
majortsuzumi 2 targets Japanese enterprise and local deployment through NTT's commercial customer network
majorSB Intuitions develops Sarashina Japanese models; SoftBank distribution and Oracle Alloy support deployment
majorVia Rakuten AI 3.0, adapts open models for Japanese use across its commerce and enterprise ecosystem
challengercotomi v3 supplies Japanese enterprise models; selection for government Gennai supports public-sector reach
majorTakane combines Japanese enterprise tuning with Cohere technology for secure customer deployments
majorELYZA unit releases ELYZA-LLM-Diffusion and supplies Japanese models to the government Gennai project
challengerBreeze 2 adapts Llama for Traditional Chinese and multimodal edge assistants; weights are publicly available
nicheFoxBrain 2.0 integrates NVIDIA Nemotron Omni for industrial reasoning and manufacturing workflows
nicheOpen CyberAgentLM3 Japanese model weights support commercial adaptation, advertising and media applications
nicheOpen Stockmark-100B Japanese-English weights support technical research and enterprise document analysis
nicheHeron open vision-language models connect Japanese visual understanding with autonomous-driving research
nicheTAIDE open models adapt Traditional Chinese language and local knowledge for Taiwan public and enterprise use
niche
Sources: taide.tw, cyberagent.co.jp, sakana.ai, tech.preferred.jp, group.ntt, global.rakuten.com
Sovereign & Open Labs
European, Indian and Middle Eastern labs offer local control, enterprise specialization or open research. State-funded institutes are included for their models rather than investable equity. The Cohere–Aleph Alpha combination is signed but remains subject to regulatory approvals.
Medium 3.5 and open-weight models serve enterprise customers; ASML and Samsung extend industrial ties
leaderCommand A+ and North target sovereign enterprise AI; Fujitsu and SAP distribute solutions
leaderPhariaAI supplies sovereign enterprise systems with Schwarz support; the Cohere combination is pending
nicheJamba Large 1.7 and Mini 2 combine long-context enterprise inference with Maestro orchestration
nicheFalcon H1R 7B provides open reasoning weights from Abu Dhabi's state-funded research program
nicheInception trains JAIS 2 with MBZUAI and Cerebras; Microsoft minority investment links the group to Azure
majorALLaM Arabic models underpin HUMAIN IQ; PIF sponsorship links local model development to sovereign cloud
emergingSarvam 105B and 30B target Indian-language reasoning; HCLTech backs commercial enterprise distribution
challengerIndian-language Krutrim-2 adapts Mistral-NeMo; the May 2026 cloud pivot limits visibility into new model releases
emergingOlmo 3 releases models and training artifacts for open research; nonprofit funding supports transparency
nicheBuilds world-model architectures with European and US backers; no generally available flagship verified
emergingDevelops learning systems with Sequoia/NVIDIA backing; no commercial flagship released by cutoff
emergingThomson adapts an open base with legal data; research weights carry noncommercial terms, distinct from commercial CoCounsel
nicheModernBERT encoders co-developed with Answer.AI support retrieval; Paradigm deploys private enterprise AI
nicheMoshi speech-to-speech and Hibiki translation models publish weights for real-time multilingual voice research
nicheCore and Flash multimodal models support enterprise reasoning; Edge targets smaller deployment environments
nicheZaya language models and Zonos speech weights target efficient open inference on alternative accelerators
nicheProject Indus trains Hindi and regional-dialect language models for Indian enterprise and public-service use
nicheParam-1 bilingual weights form part of an India-funded multilingual and multimodal research program
nicheSEA-LION open models adapt Southeast Asian languages and cultural context for local developers
niche
Sources: lighton.ai, aisingapore.org, france24.com, reuters.com, cohere.com, falconllm.tii.ae
Multimodal Model Labs
Specialist labs train image, video, speech, music and world models sold through APIs and creative tools. This segment tracks model supply rather than consumer applications; licensing rights and controllable production quality matter as much as benchmark performance.
Eleven v3 and conversational speech models supply real-time voice agents and developer APIs
leaderGen-4.5 video and GWM-1 world models serve creative production and simulation developers
majorV8.2 image generation and Niji 7 serve creators; the V8.2 editing model remains on the alpha site
majorRay3.2 video and UNI1.1 models serve creative workflows; HUMAIN supplies strategic backing
majorKling 3.0 supplies video generation APIs and creator tools within the listed platform group
majorSeedance 2.5 generates audio-video clips for Jimeng and Doubao Pro; BytePlus extends API distribution
majorFirefly Image 5 and video models integrate licensed generation into Creative Cloud workflows
majorFLUX.2 and klein models supply image generation and editing through BFL APIs and cloud partners
majorAtlas world model and Marble spatial tools turn visual inputs into interactive 3D environments
challengerv6 music generation launches with Warner, BMG and Believe partnerships for licensed creation
leaderMusic models move toward a licensed platform under UMG/Warner settlements; rollout generation is not confirmed
challengerShort-form video generation with Pika 2.5 serves creators through its web service and developer platform
challengerStable Diffusion 3.5 and Stable Audio 3.0 supply creative models; music labels join its 2026 financing
challengerVidu Q3 video and S2 Avatar serve controllable generation; Alibaba is a financial backer
challengerExpressive avatar and speech models generate enterprise training and internal-communications videos
challengerAvatar and Video Translation models generate multilingual presenter videos and lip-synchronized dubbing
challengerSonic speech models generate low-latency streaming voices for interactive agents and developer APIs
challengerOasis 3 world models simulate interactive driving and visual environments with continuous video generation
challengerMoonvalley merger adds Marey video models and licensed training-data infrastructure for creative and physical AI
challengerProprietary DoP image-to-video models control camera motion, lighting and visual effects for creators
challengerMirage avatar-video models and Captions tools generate presenter footage, speech and editing for creators
challengerWan open video models supply text-to-video and image-to-video generation for developers and cloud customers
challengerHailuo video-generation models turn prompts and images into clips through creator tools and APIs
challengerHunyuan Video and 3D model families generate motion and spatial assets through open weights and cloud tools
challenger
Sources: runwayml.com, updates.midjourney.com, suno.com, platform.stability.ai, docs.bfl.ai, blog.adobe.com
Strategic Lab Backers
Listed strategic investors exchange capital for economic exposure and commercial access to private labs. Role amounts distinguish paid investments, future commitments and dated ownership estimates. Cloud bookings, chips and model distribution can matter more than minority equity returns.
OpenAI stake was 27% at the October 2025 recap, before later dilution; $5B Anthropic equity adds Claude exposure
leaderAnthropic equity: $13B invested plus up to $20B conditional; OpenAI $50B commitment includes future tranches
leaderAnthropic financing commits $10B upfront plus $30B conditional in 2026; also holds a reported SSI minority stake
majorOpenAI $30B and Anthropic up to $10B equity commitments; minority equity also backs Mistral, Reflection and Cohere
majorOpenAI 2026 equity commitment: $20B paid in April and July, with $10B due in October; adds to earlier holdings
majorMistral equity: EUR1.3B bought about 11% in 2025; reported EUR400M September 2026 follow-on supports the stake
majorAnthropic minority equity reportedly marked near $5B in June 2026; Ventures also holds Cohere and Mistral stakes
majorAnthropic minority equity is disclosed; the $51M cost and January 2026 $2–4B stake value are Baird estimates
nicheAnthropic $100M 2023 equity investment receives a reported June 2026 follow-on; current percentage undisclosed
nicheCohere minority equity, amount undisclosed; OCI hosts Command models for enterprise database customers
nicheCohere equity participation in the 2025 financing extension; amount undisclosed, with joint accelerator optimization
nicheCohere and Mistral minority equity through the $1B AI investment fund; individual investment amounts undisclosed
nicheDirect minority equity in Anthropic, Cohere and Aleph Alpha; 2023 investment amounts remain undisclosed
nicheSeptember 2026 Mistral strategic equity investment, amount undisclosed; joint semiconductor AI development
nicheJune 2025 Scale AI investment: $14.3B for about 49% non-voting equity, supporting training-data access
majorCohere minority equity, amount undisclosed; Takane co-development adapts models for Japanese enterprise customers
nicheSarvam strategic equity investment of $150M in its 2026 financing; enterprise delivery adds Indian model distribution
nicheOwns 51% of C Holdings, which owns 50% of SB OAI Japan; the JV sells Cristal intelligence to Japanese enterprises
majorOwns 70% of the enterprise-AI joint venture with Meta; model customization and deployment serve Indian businesses
nicheMinority equity in AI21 Labs and Hugging Face through Intel Capital; individual investment amounts undisclosed
nicheMinority AI21 Labs equity through Comcast Ventures since the November 2023 Series C extension; amount undisclosed
nicheControlling equity in ELYZA since 2024 brings Japanese language-model development inside the telecom group
nicheMinority equity in Hugging Face from its 2023 Series D; NVIDIA acquisition agreement has not yet closed
niche
Sources: group.softbank, cnbc.com, tweakers.net, openai.com, anthropic.com
Other Stakes & Listed Funds
Asian minority investors and exchange-listed venture portfolios extend access to private labs. Fund weights are percentages of portfolio value, never ownership of the underlying lab. Financial holdings alone do not imply exclusive access, cloud commitments or a supply contract.
MiniMax equity about 12.6% after its July 2026 placement; minority investments also include Moonshot and Baichuan
leaderMinority equity in Moonshot and StepFun; January 2026 IPO disclosures also identify Zhipu and MiniMax stakes
majorLong-Z venture investments hold Moonshot and Zhipu minority equity; current look-through percentages undisclosed
nicheBaichuan minority equity from the July 2024 financing; investment amount and current ownership percentage undisclosed
nichePoolside $1B equity plus $6B license; SSI $5B reported equity; also backed xAI before its SpaceX merger
majorNovember 2025 Sakana minority equity investment, amount undisclosed; joint Japanese banking-model development
nicheJPY10.5B invested in Preferred Networks shares in 2017 for mobility AI; current ownership percentage undisclosed
nicheJune 30 investment portfolio: Anthropic 14.4%, OpenAI 2.6%, Databricks 1.2%; economic exposure includes SPVs
majorJune 30 NAV exposure via equity and SPVs: Anthropic >20%, OpenAI 10–20%, Databricks 10–20%; NYSE ticker VCX
majorOpenAI direct equity: $75M invested in April; June 30 portfolio weights were 11% OpenAI and 12.7% Databricks
challengerOpenAI fund interests marked $59.3M at June 30, 2026; indirect equity held through private investment vehicles
nicheAugust 2026 trust assets include 2.9% Anthropic and 4.4% ByteDance; private-equity values are manager estimates
majorFebruary 2026 minority equity investment in Sakana AI; dollar amount undisclosed, with banking AI collaboration
nicheMarch 2026 minority equity investment in Sakana AI; amount undisclosed, with industrial model co-development
nicheSeptember 16 commitment to buy JPY10B of Preferred Networks shares; shareholder approval and allotment pending
emergingListed trust holds Anthropic and OpenAI private equity; May 2026 annual report records new OpenAI investment
nicheListed fund owns Anthropic equity; June 2026 holdings report values the stake using the May Series H financing
niche
Sources: destiny.xyz, cdn.robinhood.com, mhi.com, mitsubishielectric.com, wirescreen.ai, bk.mufg.jp
Glossary
- Foundation model
- A broadly trained model reused across tasks through prompting, tools or additional training.
- Open weights
- Model parameters are available to download; the license may still restrict commercial use.
- Post-money
- Company equity valuation after new financing; it differs from money raised and enterprise value.
- Minority stake
- An economic interest that does not necessarily grant control, voting rights or exclusive model access.
- RL
- Reinforcement learning improves model behavior using rewards from verifiable tasks or human feedback.
Heat, spend, exposure and shares are editorial estimates. How to read the map ↗