Researched and adversarially fact-checked 2026-09-04. 14 labs, 13 comparable dimensions, 118 dated releases, 30 landmark papers. The interactive version of this page needs JavaScript; everything it says is below.
The labs
- Meta AI — FAIR + MSL — Frontier lab bolted onto the world's largest social distribution network; flagship Muse Spark 1.3 (2026-09); some open weights; licence: Mixed — Apache-2.0 on Muse Glimmer, bespoke Llama Community Licence on Llama, 11 distinct licences across the facebook org; 93M Hugging Face downloads a month; 60k GitHub stars.
- OpenAI — The consumer assistant that became an advertising business; flagship GPT-6 Astra (2026-09); some open weights; licence: Apache-2.0 on every model-weight release since 2025; 77M Hugging Face downloads a month; 121k GitHub stars.
- Google DeepMind — The only lab that designs its own silicon and its own science; flagship Gemini 3.8 Flash (2026-09); some open weights; licence: Mixed — Apache-2.0 on Gemma 4, bespoke Gemma Terms of Use on Gemma 1-3 and the specialist variants; 137M Hugging Face downloads a month; 199k GitHub stars.
- Anthropic — Closed weights, open protocols, and the sharpest safety paper trail; flagship Claude Fable 5.1 (2026-09); no open weights; licence: None — no model weights released under any licence; 0 Hugging Face downloads a month; 174k GitHub stars.
- DeepSeek — MIT-licensed trillion-parameter weights, and no strings at all; flagship DeepSeek-V4-Pro-0813 (2026-08); open weights are core strategy; licence: MIT; 23M Hugging Face downloads a month; 104k GitHub stars.
- Qwen (Alibaba Cloud) — The open-weight download leader, by a factor of three; flagship Qwen3.8-2.4T-A95B (2026-08); open weights are core strategy; licence: Mixed — Apache-2.0 on small and mid-size models, custom Qwen licences on the flagships; 313M Hugging Face downloads a month; 28k GitHub stars.
- Moonshot AI — Few models, all enormous, joint-first among open weights; flagship Kimi K3 (2026-06); open weights are core strategy; licence: Custom Kimi K3 License (modified-MIT style, not OSI-approved); 4.9M Hugging Face downloads a month; 11k GitHub stars.
- Z.ai (formerly Zhipu AI) — Joint-first open weights, and the only open licence in Arena's top 25; flagship GLM-5.3 (2026-08); open weights are core strategy; licence: Mixed — MIT on GLM-5.2 and GLM-5.3-Flash, custom GLM-5.3 License on the flagship; 10M Hugging Face downloads a month.
- MiniMax — Open weights with a licence that excludes half the rich world; flagship MiniMax-M3 (2026-06); some open weights; licence: Custom MiniMax community licences (not OSI-approved), one with territorial exclusions; 7.6M Hugging Face downloads a month.
- Mistral AI — Europe's champion, still living off a 2024 7B model; flagship Mistral-Large-3-675B-Instruct-2512 (2025-11); some open weights; licence: Mixed — Apache-2.0 on Mistral Large 3 and Leanstral, custom on Mistral Medium 3.5; 11M Hugging Face downloads a month; 11k GitHub stars.
- NVIDIA — Sells the compute, quantises everyone else's models; flagship NVIDIA-Nemotron-3-Ultra-550B-A55B (2026-06); open weights are core strategy; licence: OpenMDW-1.1 on Nemotron 3 (permissive, but not OSI-listed); 49M Hugging Face downloads a month.
- Microsoft Research — A large open publisher that left the LLM race; flagship Phi-4-reasoning-vision-15B (2026-01); some open weights; licence: MIT and Apache-2.0; 42M Hugging Face downloads a month.
- Allen Institute for AI (Ai2) — The only lab that ships the training data with the weights; flagship Olmo-3-7B-Instruct (2025-11); open weights are core strategy; licence: Apache-2.0; 6.4M Hugging Face downloads a month; 6.7k GitHub stars.
- xAI / SpaceXAI — Frontier-adjacent scores, almost no public footprint; flagship Grok 4.6; some open weights; licence: Unverified; 10k Hugging Face downloads a month.
The dimensions compared
- Does the lab publish model weights at all? — Is releasing weights the strategy, a side-channel, or something the lab has never done? 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? 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? 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? 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? 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? 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? 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? 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? 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? 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 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? 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? 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.
Headline findings
- Two-thirds of Meta's public models forbid commercial use — and one-ninth of its downloads do. Measured across Meta's 'facebook' Hugging Face org on 2026-09-04: 1,554 of 2,359 models (66%) carry a non-commercial licence, but those models account for only 11.4% of the org's trailing-30-day download volume, while 51.3% of volume is OSI-permissive. The count says 'less open than you think'; the volume says 'the parts you actually use are open'. Both are true; printing one without the other is PR in whichever direction it is trimmed. Bucketing rule: OSI-permissive counts apache-2.0, mit, bsd-2/3-clause and cc0-1.0 only; undeclared-licence repos are counted in neither bucket — which matters, because Meta's second most-downloaded model, Contriever, declares no licence at all. source
- Every Llama repo makes you ask permission first. Qwen, Gemma 4, gpt-oss and Meta's own Muse Glimmer do not. All 70 repos in the meta-llama org read gated='manual' in the Hugging Face API — 100% of that org's 25,761,588 trailing-30-day downloads require requesting access and waiting for Meta to approve it. Qwen (465 repos), openai (39) and meta-models (4) gate nothing at all; Google gates 352 of 1,133 repos, 10.65% of its volume. 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 own download numbers look smaller, not larger. source
- Meta gave the industry PyTorch. It is not in the room where the agent standard is being written. Meta transferred PyTorch (102,743 GitHub stars) to the newly-formed PyTorch Foundation under the Linux Foundation on 12 September 2022 — the largest infrastructure donation of the era, and it predates the current model race. The agent era's equivalent standard, the Model Context Protocol, was created by Anthropic and donated to the Linux Foundation's Agentic AI Foundation on 9 December 2025; its servers repo has 90,061 stars. The Foundation's Platinum members are AWS, Anthropic, Block, Bloomberg, Cloudflare, Google, Microsoft and OpenAI. Meta appears at no tier. Absence from a member list is not a statement of intent — report the fact, do not infer a motive. source
- Eleven of Meta's twelve most-downloaded models are not chat LLMs — and neither are OpenAI's top three. Meta's top twelve by trailing-30-day downloads on 2026-09-04: opt-125m 9.68M (the only generative LLM), Contriever 7.74M, DINOv2-small 4.37M, wav2vec2-base 3.12M, BART-MNLI 3.08M, DINOv2-base 2.84M, w2v-BERT 2.0 2.25M, SAM 3 2.09M, MusicGen 1.88M, ESMFold 1.75M, NLLB-200 1.44M, ESM-2 1.40M. At OpenAI, CLIP (19.94M), Whisper large-v3-turbo (6.84M) and CLIP-large (6.72M) all out-download gpt-oss-20b (6.21M). Every standard comparison axis — context window, MMLU, Elo, tokens per second — is defined only for language models, so the leaderboard everyone quotes covers a minority of what these labs ship. source
- Meta reports Muse Glimmer at 76.0 on SWE-Bench Verified. An independent index scores the same model 35, mid-pack among its peers. Meta's own model card reports MCP Atlas 75.5, SWE-Bench Pro 51.2 and SWE-Bench Verified 76.0 — and in the same table concedes losing SWE-Bench Verified to Qwen3.6-27B at 77.2, and GPQA Diamond to both comparators. Across the card's 22 scored rows Muse Glimmer leads 12 and trails 10. Artificial Analysis, measuring independently, puts its Intelligence Index at 35: behind Qwen3.6 27B and Ling 3.0 Flash (both 38) and Kimi K2.5 (36), ahead of Gemma 4 31B (30). It also measures an 82% hallucination rate against 49% for Qwen3.6 27B. Two clearly-labelled columns, never one: the numbers measure different things and neither is a lie. Do NOT imply Meta inflated SWE-Bench Verified — Meta's own table shows a rival ahead on that exact benchmark. source
- Meta's open model ranks 11th of 14 on the independent open-weight board — while its closed model ranks 5th on Arena. Artificial Analysis's open-weight Intelligence Index (v4.1.1), read 2026-09-04, is led by Moonshot's Kimi K3 and Z.ai's GLM-5.3 at 60, then Qwen3.8 2.4T A95B (58), GLM-5.3-Flash (57), DeepSeek V4 Pro (53). Meta's Muse Glimmer sits at 35, eleventh. On Arena's blind human-preference board the same week, Meta's PROPRIETARY muse-spark-1.2 (xHigh) ranked 5th at 1499 Elo — above every Chinese lab's entry, and the highest-ranked non-Anthropic model on the board. The honest sentence is therefore: Meta trails on what it open-sources, not on what it builds. Arena's Elo intervals overlap heavily across ranks 5-15, so treat the exact rank as provisional. source
- Three labs, three user numbers, three different units — and one of them is over a year old. OpenAI disclosed more than 900 million weekly active users on 27 February 2026, and stated over 1 billion weekly actives in August 2026. Google announced the Gemini app had surpassed 1 billion MONTHLY users on 11 August 2026 — its own wording is 'monthly users', not monthly actives. Meta's most recent official Meta AI figure is 1 billion monthly actives from May 2025, unrefreshed since; 2026 aggregator figures range from 500M to 1.5bn and contradict each other, and Meta's Q2 2026 press release discloses no assistant user number at all. Weekly actives are strictly smaller than monthly actives for the same product, so OpenAI's smaller-looking number is probably the larger audience. Do not draw bars from these three figures under any circumstances. source
- The only audited numbers on the page: Meta guides $130-145bn of 2026 capex, and its free cash flow fell 91% in a single year. Verbatim from Meta's Q2 2026 release (29 July 2026): 'We anticipate 2026 capital expenditures, including principal payments on finance leases, to be in the range of $130-145 billion, narrowed from our prior outlook of $125-145 billion.' In the same filing, Q2 2026 free cash flow was $784 million against $8,549 million a year earlier — a 90.8% fall — with operating margin down from 43% to 31%. Alphabet's Q2 2026 free cash flow was NEGATIVE by roughly $5.8bn as $44.9bn of capex exceeded $39.1bn of operating cash flow. Alphabet's reported $195-205bn full-year guidance is call reporting, not a filed figure, and could not be reopened from any reachable source. Meta reports no AI revenue segment at all — putting a Meta 'AI revenue' number beside OpenAI's or Anthropic's run rates would be fabricating one. source
- Meta is the only frontier lab whose AI ships on people's faces — and the only one whose custom silicon nobody else can use. In August 2026 Meta gave 15,000 Ray-Ban Meta glasses free to every blind and visually impaired adult supported by Vision Ireland, built with EssilorLuxottica, for text reading, object identification, live translation and hands-free calls. OpenAI's first device is targeted for H2 2026 and did not exist at retail as of 2026-09-04. In the same year Meta announced four MTIA generations in a single day (11 March 2026) with hundreds of thousands of chips deployed for inference — none of them available outside Meta, while Google both trains its frontier models on its own TPUs and is reported to have begun selling them for third-party data centres. Lead and trail in a single row. Do not extrapolate the 15,000 figure into market share; it is a donation, not a shipment number. source
- No new Llama has shipped since April 2025 — but Meta never stopped publishing open weights. The newest upload of any kind to Meta's meta-llama Hugging Face org is Llama-Prompt-Guard-2-86M, created 28 April 2025; the Llama 4 model repos date to 1-2 April 2025. llama.com now redirects to Meta's general AI developer site, Meta's own model catalogue lists no Llama 5, and the llama-stack GitHub repository has been transferred out of the meta-llama org. Meta has never announced a cancellation — asked directly, a spokesperson said only 'Our current Llama models will continue to be available as open source', addressing availability while dodging development. But the separate 'facebook' org kept shipping open weights throughout: Perception Encoder (April 2025), MIT-licensed V-JEPA 2 weights (May 2025), map-anything (September 2025), omniASR (November 2025), MobileMoE (June 2026) and MoEViE (August 2026). 'Meta stopped releasing open models' is false; 'Meta stopped releasing Llama models' is true. Word it precisely. source
Where Meta stands
Leads: Shared infrastructure, on cumulative contribution: PyTorch (102,743 stars, given to the Linux Foundation in 2022), FAISS under MIT (40,852 stars) and ONNX, co-created with Microsoft in 2017 and handed to LF AI & Data in 2019.; The largest public model catalogue measured of any frontier lab: 2,433 models across three Hugging Face orgs, against Google's 1,139 and OpenAI's 46.; Breadth outside language: eleven of Meta's twelve most-downloaded models are not generative chat LLMs, spanning segmentation, vision, speech, proteins, music and translation.; Deployed non-benchmark science: SAM 3 and DINOv3 in the DOE Genesis Mission at Lawrence Berkeley National Laboratory, and DINOv3 plus Segment Anything running on-device in an ARPA-H-funded assistive-robotics programme — both per Meta's own account.; Consumer hardware: the only frontier lab whose AI ships on a wearable people actually buy, and the only one to announce four custom-silicon generations in a single day.; Cost per unit of measured intelligence at the frontier: Artificial Analysis records Muse Spark 1.3 (xhigh) as having the lowest cost per task of any model scoring 59+ on its Intelligence Index, at $0.55 per task.; Release cadence at the frontier: four Muse Spark versions between April and September 2026..
Trails: Open-weight capability: Muse Glimmer scores 35 on Artificial Analysis's open-weight Intelligence Index — eleventh of fourteen, roughly 25 points behind Kimi K3 and GLM-5.3.; Open-weight reach: third on Hugging Face download volume at 92.6M in 30 days, behind Qwen's 312.6M and Google's 136.8M — and Alibaba leads on cumulative all-time downloads too.; Licence friction on the legacy line: 100% of Llama repos are gated behind manual approval, and the Llama 4 Community Licence's 700M-MAU clause, 'Built with Llama' requirement and name-prefix rule keep it outside the OSI definition.; Standards participation: absent at every tier from the Agentic AI Foundation, whose Platinum members include every other major lab on this page.; Dataset publishing: 123 public Hugging Face datasets, behind AI2's 1,286, NVIDIA's 311 and Microsoft's 111.; The Llama line itself: nothing new since April 2025, a flagship (Behemoth) previewed and never released, and no commitment from Meta to develop it further.; Disclosure: no AI revenue segment, no refreshed assistant user figure since May 2025, and a free cash flow that fell 90.8% year-over-year in Q2 2026.; Measured safety on its own open model: Meta's card shows Muse Glimmer with higher privacy-violation and prompt-injection-success rates than Google's comparable Gemma 4 31B..
Contested: Whether Muse Glimmer is 'open source'. The Apache-2.0 licence is OSI-approved, but Meta released neither training data nor training code, and Artificial Analysis scores its Openness Index at 44 — mid-pack, level with DeepSeek V4 Flash, GLM-5.2 and Ling 3.0 Flash. 'Open weights under an OSI-approved licence' is the accurate phrase; Meta and much of the press say 'open source'.; Whether Llama is finished. No new model since April 2025, llama.com redirecting away, no Llama 5 in Meta's own catalogue and llama-stack transferred out of the org all point one way — but Meta has never announced a cancellation, and its spokesperson's answer addressed availability rather than development.; The Muse Spark open-weights promise. Zuckerberg promised it in an X post on 2 September 2026 with no date; The Register records no version number while Wikipedia says 1.2 is the version planned. Two days old at the time of writing — outstanding, not overdue.; Meta's May 2026 layoffs. Fortune reports 8,000 and a 21 May start; The Register reports ~7,800 and 20 May. Meta's own headcount disclosures show a net quarterly fall of 2,514 (77,986 at 31 March to 75,472 at 30 June), which the company has not reconciled.; Whether Meta rents Google TPUs. Reported by The Information and relayed in February 2026; neither company has confirmed it on the record.; Muse Image's #2 Arena ranking for text-to-image. This is Meta reporting a third-party leaderboard position, self-dated 5 July 2026, which this research could not independently confirm — and Meta's April 2025 LMArena episode, in which it submitted a specially-tuned non-public Llama 4 Maverick variant scoring 1417 Elo, is the reason to check.; Alphabet's $195-205bn 2026 capex guidance and Alphabet's negative free cash flow being a first. The negative figure is verified from the SEC exhibit; the guidance is call reporting only, and the 'first on record' superlative could not be established..
On the axis everyone measures — one closed frontier model against another — Meta is genuinely competitive and no longer the story: its Muse Spark line sits fifth on blind human preference and sixth on the independent intelligence index, behind Anthropic and level with OpenAI, and it is the cheapest way to buy intelligence at that level. On the axis Meta made its name on, it has been overtaken: Alibaba's Qwen out-downloads Meta's whole Hugging Face estate more than three to one, and Meta's open model ranks eleventh of fourteen on the independent open-weight board, behind four Chinese labs, an Abu Dhabi institute, NVIDIA and a US startup. The licence picture has inverted in a way nobody predicted: Meta and Google both moved to Apache 2.0 in 2026 while Alibaba and Moonshot adopted the scale-cap licences Meta invented — and Meta's own catalogue now runs a permissive licence and a restrictive one side by side, with every Llama repo still gated. The dimension where Meta clearly leads is the one no leaderboard shows: it gave the field PyTorch, FAISS and half of ONNX, its most-used open models are vision and speech rather than chat, and those models are running in national laboratories and on wheelchairs — while it sits out the standards body writing the agent era's protocols. Read as a whole, this is a lab that is spending more than almost anyone, disclosing less than almost anyone, and whose ranking changes completely depending on which axis you pick — which is why this page is a set of dimensions and not a scoreboard.
How to distrust this page
- Open weights means you can download it. Open source means you can use it. Meta's own catalogue contains models that are one, the other, both and neither.
- Two-thirds of Meta's public models forbid commercial use. One-ninth of its downloads do. Both are true; neither alone is honest.
- Every Llama repo makes you ask permission first. Qwen, Gemma 4, gpt-oss and Muse Glimmer do not. That difference never appears in an 'open models' comparison.
- A download is a CI job, a quantisation, a Docker layer or a curious human. Most are not a user.
- Never put a lab's own score next to someone else's measurement. Both numbers can be real and still not be the same kind of thing.
- Arena measures which answer people prefer. Artificial Analysis measures which model scores higher. They disagree, and 'the best model' is not a fact until you say which one you mean.
- OpenAI reports weekly users, Google reports monthly users, and Meta's newest official figure is from May 2025. Put those three in one bar chart and you have drawn a picture of nothing.
- '117B parameters with 5.1B active.' A model's headline size and its working size can differ by more than twenty times.
- The lab that popularised a project is often not the lab that built it: Triton's own licence credits a graduate student, and vLLM came out of UC Berkeley.
- A safety framework is a document, not a measurement. METR indexes twelve of these policies and says plainly that indexing them 'should not be considered an endorsement of their substance.'
- A lab with no open weights has no download number, and that is a strategy, not a weakness.
- Nobody is vertically integrated except Google. Meta designs its own chips and is reported to rent Google's; Anthropic runs on Google, Amazon and NVIDIA silicon at once.
- Capex is company-wide, no filer breaks out AI capex, fiscal years do not align, and OpenAI, Anthropic and xAI file nothing at all.
- Citation counts mostly measure how old a paper is. 'Attention Is All You Need' will outrank any 2025 paper indefinitely.
- Hugging Face downloads and likes, GitHub stars, Arena scores and citation counts are live counters. Every one here is a point-in-time reading on its stated date.