Every figure on the Frontier comparison

The interactive comparison draws charts. This page is the same data as plain tables — every dimension's per-lab value, every dated release and every landmark paper, each with the note that qualifies it, the date it was read and a link to where it came from. It needs no JavaScript and it prints.

14 labs · 13 dimensions · 118 releases · 30 papers. Compiled 2026-09-04. Back to the comparison.

The labs — 14 organisations

The figures on each lab card, in the order the comparison shows them. Ordered by trailing-30-day Hugging Face downloads.

Qwen (Alibaba Cloud): 322 of Qwen's 465 repos are Apache-2.0 and none is gated. But its flagships carry scale gates structurally similar to Meta's old 700M-MAU clause — and the Qwen Community License 1.0 on Qwen3.8-Flash-Next is stricter than the Max licence on one clause, having no revenue floor at all on the Model-as-a-Service requirement. Google DeepMind: Gemma 4 (April 2026) moved to Apache 2.0 with no MAU cap. The older Gemma Terms, revised the day before, explicitly exclude Gemma 4 and retain a binding Prohibited Use Policy plus a right for Google to 'restrict (remotely or otherwise) usage' — clauses no OSI licence has. Whether Gemma 4's separately-published Prohibited Use policy is incorporated into the Apache grant was not established from the licence text. Meta AI — FAIR + MSL: The Llama 4 Community Licence requires a separate discretionary licence from Meta above 700M monthly active users, mandates 'Built with Llama' attribution and a 'Llama' name prefix on derivatives — so it is not OSI open source. Muse Glimmer (August 2026) is plain Apache-2.0 with none of those clauses. Both are live at the same time. OpenAI: gpt-oss-120b/20b, gpt-oss-safeguard, privacy-filter and circuit-sparsity all carry license:apache-2.0 — OSI-approved, no MAU cap, no acceptable-use addendum, no gating. Still open weights and not open source: no training data or code. And OpenAI's single most-downloaded artefact, CLIP, declares no licence on Hugging Face at all. NVIDIA: OpenMDW v1.1 requires retaining the agreement and origin notices on distribution, places third-party rights clearance on the licensee, and terminates the grant if the licensee brings patent or copyright litigation over the model materials. It imposes no restrictions on model outputs. Notable that NVIDIA chose it over Apache-2.0. Microsoft Research: Microsoft's open releases are genuinely permissive — Phi and Fara1.5-27B under MIT, Mage-VL under Apache-2.0 — which makes the absence of a current frontier open LLM, rather than the licence, the story. DeepSeek: MIT is OSI-approved with no usage restrictions, no MAU cap, no revenue gate and no attribution obligation beyond the copyright notice. DeepSeek is the most permissive of the large Chinese labs, and the only one shipping MIT at the flagship tier — Qwen, Moonshot and Z.ai all use custom licences on their flagships. Mistral AI: Mistral's open-weight posture is no longer uniform: Mistral Large 3 and Leanstral-1.5-119B-A6B are Apache-2.0, while Mistral-Medium-3.5-128B is tagged license:other with no licence name in the metadata, so its exact terms are unverified. Z.ai (formerly Zhipu AI): The GLM-5.3 License is MIT-shaped but conditional: a Model-as-a-Service operator with affiliates above US$10bn of aggregate 12-month revenue must pass a Z.AI security review, scoped by Z.AI, before any commercial use. MiniMax: The most geographically restrictive licence in this survey: the MiniMax H3 Community License grants rights only in the 'Applicable Territory', defined as worldwide excluding the European Union, the United Kingdom, the Republic of Korea and the United States of America. It also imposes a separate acceptable-use policy. Allen Institute for AI (Ai2): Apache-2.0 on the weights, with training data and code released alongside — the closest thing in this survey to open source in the full sense rather than open weights. The model card adds a non-licence note that the model is intended for research and educational use in accordance with Ai2's Responsible Use Guidelines. Moonshot AI: MIT-shaped but conditional: a separate Moonshot agreement is required for Model-as-a-Service businesses above US$20M of aggregate 12-month revenue, and attribution is required above 100M MAU or US$20M monthly revenue. Internal use and access via Moonshot's own products are carved out. xAI / SpaceXAI: The xai-org Hugging Face account holds 2 public models drawing 10,131 downloads in the trailing 30 days. This pass did not verify which models those are, their licences, or whether any current Grok weights have been released. Anthropic: There is no Anthropic weight licence to compare, because there is no Anthropic weight release. Its Hugging Face org is verified and hosts 0 models and 14 datasets. This is a licence-category absence, not a restrictive licence. Anthropic's stated position (27 July 2026) is that it 'has never advocated for a ban on open-weights models' and that capability-safe open-weights models are 'a public good'.

Does the lab publish model weights at all?

Is releasing weights the strategy, a side-channel, or something the lab has never done? Read 2026-09-04. Measured in posture (none / some / core-strategy).

A lab with no open weights has no download number, and that is a strategy, not a weakness. Anthropic's absence from every open-weight chart on this page says nothing about the quality of its models — and Anthropic's own published position is that it has never advocated banning open-weight models.

Hugging Face downloads, trailing 30 days

Whose open weights are actually being pulled right now? Read 2026-09-04. Measured in downloads in the trailing 30 days.

A download is a CI job, a quantisation, a Docker layer or a curious human — most are not a user. Org totals are dominated by small legacy encoders rather than flagship LLMs (one 2020 model, google/electra-base-discriminator, is 42% of Google's total; facebook/opt-125m and facebook/contriever lead Meta's). Counts credit the namespace hosting the files, so NVIDIA's biggest streams are other labs' models and Meta's own Muse Glimmer repo is out-downloaded by a community GGUF mirror. Gated repos suppress casual downloads, which cuts against Meta. And a zero here can mean a deliberate closed-weights strategy.

Public models on Hugging Face

How large is the lab's published catalogue? Read 2026-09-04. Measured in public model repositories.

Catalogue size is a research-archive statistic, not a product statistic. Two-thirds of Meta's 2,359-model facebook org carries a non-commercial licence, and most of AI2's 970 are small research checkpoints. A large number here says a lab publishes a lot, not that it publishes well.

Public datasets on Hugging Face

Who releases the data, not just the weights? Read 2026-09-04. Measured in public dataset repositories.

This is the dimension on which every large lab, Meta included, trails a small nonprofit. Dataset count is also a crude proxy: one well-documented pre-training corpus matters more than fifty evaluation splits, and this count does not weight them.

What the licence on the flagship open model actually permits

Can you use it in a commercial product without asking anyone? Read 2026-09-04. Measured in this page's four-level coding of the licence text (3 = OSI-approved permissive, 2 = permissive-but-not-OSI, 1 = custom licence with scale, revenue or territory gates, 0 = no weights released).

The 0-3 scale is THIS PAGE'S coding of licence texts we read, not a figure any source publishes — the licence names and clauses under it are primary-sourced, the number is editorial. And an OSI-approved licence on the weights is still not 'open source' in the full sense while training data and code are withheld: that is true of Muse Glimmer, Gemma 4 and gpt-oss alike. Only AI2's OLMo ships the whole pipeline.

How much of the lab's open-weight volume needs permission first

Can you just download it, or must you ask and wait for manual approval? Read 2026-09-04. Measured in percentage of trailing-30-day download volume behind a manual Hugging Face access gate.

Gating is a real barrier that no 'open models' table has ever shown, and it is a single boolean field in a public API. It cuts both ways: it suppresses casual downloads, so it makes Meta's download numbers look SMALLER, not larger. Meta's figure here is derived arithmetic over three orgs (6.46% of the facebook org's 65.6M, 100% of meta-llama's 25.76M, 0% of meta-models' 1.20M) rather than a single published number.

Who built the commons everyone else runs on

Which lab gave away a piece of infrastructure the whole field now depends on — and then gave up control of it? Read 2026-09-04. Measured in GitHub stars on the lab's flagship donated / foundation-hosted infrastructure project.

Stars are cumulative and never decay, and they measure developer attention rather than usage. More important: the lab that popularised a project is often not the lab that built it. vLLM (90,919 stars), the inference engine most of the open ecosystem runs on, came out of UC Berkeley's Sky Computing Lab, not any frontier lab; Ray also came from Berkeley; and Triton's own LICENSE credits Philippe Tillet for 2018-2020 before OpenAI for 2020-2022. Infrastructure scoreboards that assign every project to a corporate logo systematically overstate labs and erase universities.

Membership of the Agentic AI Foundation

Who is in the room where the agent-era standard is being written? Read 2025-12-09. Measured in membership tier as listed by the Linux Foundation.

Absence from a member list is not a statement of intent. Report the fact; do not infer a motive. The list is as of the foundation's formation announcement and may have changed since.

Citations on this page's sample of landmark papers

Whose published research is the field standing on? Read 2026-09-04. Measured in summed Semantic Scholar citations across a 30-paper hand-picked sample.

This is a SUM OVER A HAND-PICKED SAMPLE of 30 landmark papers (at most a third from Meta), not a measure of total research output — a lab that publishes little appears rarely, which is not a judgement of its work. And citation counts are dominated by paper age: 'Attention Is All You Need' (2017) and BERT (2019) will outrank any 2025 paper indefinitely. Compare within publication-year cohorts, or the chart just re-discovers that old papers are old. Semantic Scholar also holds duplicate records (DeepSeek-V3) and returned nothing at all for the original Gemini 1.0 paper.

Frontier safety policy: what is written down, and when it was last updated

What has each lab committed to in public, at what specificity, and how old is the commitment? Read 2026-09-04. Measured in publication date of the most recent version, as indexed by METR.

METR maintains this index and states plainly that 'our indexing these documents should not be considered an endorsement of their substance.' Comparing policies compares what companies have promised, in what detail and how recently — not what they do. Recency is not goodness. And a common false contrast should be avoided: Meta's tiers (Critical / High / Moderate-or-lower) and Anthropic's ASL levels are both named tiers rather than numeric scores; the real difference is in how specifically the capability triggers are defined, not in whether thresholds exist.

Capital expenditure — the only audited numbers on the page

Who is actually spending the money, and can you even see it? Read 2026-09-04. Measured in USD billions, purchases of property and equipment, most recent audited fiscal year (period stated per row).

Read every row's period before comparing: fiscal years DO NOT align. Meta and Alphabet are calendar-year filers (FY2025); Microsoft's FY2026 ran July 2025 to June 2026; NVIDIA's FY2026 ended 25 January 2026. Capex is company-wide and no filer breaks out AI capex. NVIDIA is a supplier, not a buyer — it sells the compute — so its tiny figure is not a signal of restraint. OpenAI, Anthropic and xAI file nothing at all, so their cells are empty by necessity. Meta's own headline guidance ($130-145bn for 2026) is on a different basis from the GAAP line here, because it includes principal payments on finance leases — which is why press quotes Meta's 2025 capex as ~$72.2bn while the 10-K line reads $69.691bn. Pick one basis and label it.

What the lab ships that is not a chatbot

How much of the work does an LLM leaderboard structurally fail to see? Read 2026-09-04. Measured in inventory of published non-LLM model families (qualitative).

This row is deliberately not a number, because there is no honest one: the labs publish in different formats and count different things. What IS measured is the download shape — eleven of Meta's twelve most-downloaded Hugging Face models are not generative chat LLMs, and neither are OpenAI's top three, which all out-download gpt-oss-20b. Every standard comparison axis (context window, MMLU, Elo, tokens per second) is defined only for language models, so breadth is invisible by construction rather than by evidence.

Silicon, and things people actually wear

Who controls their own compute, and whose AI exists as a physical object? Read 2026-09-04. Measured in hardware position (qualitative).

Hardware comparisons imply an independence the supply chain does not support. Meta designs its own chips and is reported (by The Information, unconfirmed by either company) to rent Google's; Anthropic runs on Google, Amazon and NVIDIA silicon simultaneously; OpenAI's compute is contracted across Nvidia, Oracle, AMD, Broadcom and AWS. Nobody here is vertically integrated except Google. Note also that Meta's newsroom is silent on whether MTIA could be offered externally — say 'Meta has not offered MTIA outside its own infrastructure', not that it cannot be.

The release race — 118 dated releases

Newest first.

What the field builds on — 30 landmark papers

Older papers have had longer to accumulate, so this ranks foundations, not current standing.