← Gautam Parab

The Biggest Language Model Is 2.8 Trillion Parameters. 83 Percent of Downloads Go to Models Under One Billion.

Ask “how big is the biggest AI model” and the honest answer splits into two questions that don’t share evidence. Among labs that publish a number, the largest is Moonshot AI’s Kimi K3: 2.8 trillion total parameters, a mixture-of-experts model that activates 16 of 896 experts per token for a working set of 104 billion parameters, with a context window of just over a million tokens. Among labs that don’t publish a number, the largest model is unknown, because OpenAI, Anthropic, and Google DeepMind have never disclosed a parameter count for GPT-6 Astra, any Claude model, or any Gemini model. Every “Claude is probably 5 trillion parameters” figure you’ve seen traces back to an offhand comment, not a filing or a technical report. The three labs whose models most people actually mean when they say “biggest AI” are the three labs about which the question is, technically, unanswerable.

What “biggest” and “smallest” actually mean now

Kimi K3 is the largest open model by both total and active parameters at once, which is unusual: most large mixture-of-experts systems post a huge total figure while keeping the active count comparatively modest. DeepSeek’s V4-Pro model, released in April, carries 1.6 trillion total parameters but activates only 49 billion per token. Meta’s Llama 4 Maverick, disclosed the year before, goes further in the other direction: 400 billion total parameters routed through 128 experts, of which only 17 billion activate on any given pass. A model’s total parameter count tells you how much it knows how to route between. Its active count tells you how much compute you’re actually paying for on a given request. Reporting only the first number, which is what most coverage does, is a little like describing a company by its total headcount when the number that determines your bill is how many employees show up to any one meeting.

Total parameters vs. active parameters, log scale, six models A paired-bar comparison on a logarithmic scale. Kimi K3: 2.8 trillion total, 104 billion active, the highest active-parameter count of any model shown despite still being a small fraction of its own total. DeepSeek-V4-Pro: 1.6 trillion total, 49 billion active. Llama 4 Maverick: 400 billion total, 17 billion active. Apple AFM 3 Core Advanced: 20 billion total, 1 to 4 billion active depending on the request. Gemma 4 E2B: 5.1 billion total, 2.3 billion effective through a Per-Layer Embeddings architecture. GPT-6 Astra, Claude, and Gemini are shown as a separate row marked undisclosed, since none of the three labs has ever published a parameter count for any current model. TOTAL VS. ACTIVE PARAMETERS, LOG SCALE · SEP 2026 Kimi K3 DeepSeek-V4-Pro Llama 4 Maverick Apple AFM 3 Core Advanced Gemma 4 E2B Kimi K3 total: 2.8T parameters Kimi K3 active: 104B parameters DeepSeek-V4-Pro total: 1.6T parameters DeepSeek-V4-Pro active: 49B parameters Llama 4 Maverick total: 400B parameters Llama 4 Maverick active: 17B parameters Apple AFM 3 Core Advanced total: 20B parameters Apple AFM 3 Core Advanced active: 1-4B parameters Gemma 4 E2B total: 5.1B parameters Gemma 4 E2B effective: 2.3B parameters 2.8T / 104B 1.6T / 49B 400B / 17B 20B / 1-4B 5.1B / 2.3B eff. GPT-6 Astra, every Claude model, every Gemini model No parameter count has ever been published by OpenAI, Anthropic, or Google DeepMind. Dark bar: total parameters. Light bar: active (or effective) parameters. Both on the same log scale.
Total parameters measure how much a model knows how to route between. Active parameters measure what you're actually paying compute for. Reporting only the first number obscures the second.

Smallest is just as slippery, in the other direction. Google’s Gemma 4 E2B, released in April, lists 5.1 billion total parameters but runs at an effective 2.3 billion through an architecture Google calls Per-Layer Embeddings, which keeps most of the parameter mass off the accelerator at inference time. Apple’s AFM 3 Core, shipped in June as the on-device model inside its own products, is a plain 3-billion-parameter dense model with no such trick. Apple’s larger sibling, AFM 3 Core Advanced, is a 20-billion mixture-of-experts model that activates somewhere between 1 and 4 billion parameters depending on what you ask it, which means the same model can report two different “sizes” for two different requests in the same session. A parameter count used to be one number. For a meaningful share of current models, on both the top and bottom end, it’s now the headline figure for an answer that’s actually a range.

The distribution nobody headlines

Here’s the number that should reframe the other two. Hugging Face’s own mid-year accounting of its Hub, published as “State of Open Models: Summer 2026,” states plainly: “among models that declare a parameter count, those under 1B take 83% of all-time downloads and everything above 100B takes 1%.” Restricting to downloads accumulated in 2026 alone changes almost nothing: “3% of the volume goes to models above 70B.” A separate finding in the same post is more extreme still: “roughly 85.6% of models have fewer than 200 lifetime downloads, and 1.5% of repositories account for 99.2% of all downloads.” Hugging Face caveats its own number: “Downloads indicate usage within the Hub ecosystem, but they do not capture API usage, private deployments, or models distributed through other channels.” Most real production traffic to GPT-6 Astra or Claude never touches a Hugging Face download counter, because those models aren’t downloadable at all. So this isn’t a market-wide median. It’s the clearest measurement anyone has published of what happens when a huge, open population of users can pick any model size they want, and it says the population overwhelmingly picks small.

Hugging Face Hub downloads by declared model size, 2026 A single horizontal bar divided into three segments by share of all-time Hub downloads. Under 1 billion parameters: 83 percent of downloads. 1 billion to 100 billion parameters: 16 percent of downloads, the remainder. Above 100 billion parameters: 1 percent of downloads. A separate annotation notes that restricting to 2026-only downloads changes the top end only slightly, to 3 percent for models above 70 billion parameters. HUGGING FACE HUB, ALL-TIME DOWNLOADS BY DECLARED SIZE Under 1B parameters: 83% of all-time downloads 1B to 100B parameters: 16% of all-time downloads Above 100B parameters: 1% of all-time downloads 83% · under 1B parameters 16% · 1B-100B 1% · over 100B 85.6% of models on the Hub have under 200 lifetime downloads 1.5% of repositories account for 99.2% of all downloads. Hub downloads measure Hub-ecosystem usage only, not API traffic or private deployments; Hugging Face states this caveat itself.
Restricting to 2026-only downloads barely moves the top end: 3 percent of the year's volume still goes to models above 70 billion parameters.

That skew doesn’t fully hold once you leave Hugging Face and look at where paid inference traffic actually goes. I checked OpenRouter’s own leaderboard directly: it had Claude Fable 5.1, Qwen3.8 Max, GPT-6 Astra, Claude Opus 5, Claude Fable 5, and GPT-5.6 Sol filling out its top ranks by token volume in mid-September: frontier-class models, not small ones. But the composition of that traffic has shifted hard toward cost. OpenRouter’s weekly volume hit roughly 25 trillion tokens by late May, a fivefold jump in six months. Secondary trackers, not OpenRouter’s own published figures, put Chinese-origin models (mostly cheaper, often sparser mixture-of-experts systems) at under 2 percent of that traffic in late 2024 and somewhere near 45 percent by this spring; I couldn’t find an OpenRouter source stating both ends of that comparison itself, so treat the trend as directionally likely and the exact numbers as unconfirmed. The frontier names still top the leaderboard. The money is increasingly flowing to whichever frontier-adjacent model costs the least per token to run, which is a different competition than the one “biggest model” headlines describe.

What actually changes with size, task by task

The clean academic version of that same story comes from a 2025 IEEE study comparing small language models under 250 million parameters against full-scale frontier models (Mistral 7B, Qwen3-235B, and DeepSeek-R1 at 671 billion) across six tasks. On four of the six, the tiny fine-tuned models matched or beat the giants outright: a 14-million-parameter ELECTRA model hit 95 percent accuracy on sentiment classification against Mistral 7B’s 68 percent; a 250-million-parameter FLAN-T5 scored 88 percent on natural language inference against Mistral 7B’s 62 percent; a 110-million-parameter CodeParrot model beat DeepSeek-R1 on a code-summarization metric by more than double. The energy cost difference for these wins ran as high as 13,000 times less carbon dioxide per query. But on the two tasks that require reasoning depth (chain-of-thought problems and code generation from scratch), the gap reversed completely: DeepSeek-R1 scored 65.9 percent on HumanEval against a 150-million-parameter model’s 12.8 percent. Size buys you almost nothing on tasks a small, well-tuned model already handles, and almost everything on tasks that require holding a long reasoning chain together. A single parameter count can’t tell you which regime you’re in.

A commercial example from April makes the same point more starkly. Alibaba’s Qwen team released a 27-billion-parameter dense model, Qwen3.6-27B, and reported it beating their own much larger Qwen3.5-397B-A17B (a mixture-of-experts model with nearly fifteen times the total parameter count) on every agentic coding benchmark they ran it against.

A 27-billion-parameter dense model outscoring a 397-billion-parameter mixture-of-experts model A dumbbell chart across four benchmarks comparing Qwen3.6-27B, a 27 billion parameter dense model, against Qwen3.5-397B-A17B, a 397 billion total parameter mixture-of-experts model with 17 billion active. SWE-bench Verified: 77.2 versus 76.2. SWE-bench Pro: 53.5 versus 50.9. Terminal-Bench 2.0: 59.3 versus 52.5. SkillsBench: 48.2 versus 30.0, the largest gap. In every case the smaller dense model scores higher. These are vendor-reported results using Qwen's own evaluation harness, not independently replicated benchmarks. QWEN3.6-27B (DENSE) VS. QWEN3.5-397B-A17B (MOE) · VENDOR-REPORTED SWE-bench Verified SWE-bench Pro Terminal-Bench 2.0 SkillsBench Qwen3.5-397B-A17B: 76.2 Qwen3.6-27B: 77.2 Qwen3.5-397B-A17B: 50.9 Qwen3.6-27B: 53.5 Qwen3.5-397B-A17B: 52.5 Qwen3.6-27B: 59.3 Qwen3.5-397B-A17B: 30.0 Qwen3.6-27B: 48.2 77.2 53.5 59.3 48.2 76.2 50.9 52.5 30.0 Dark dot: Qwen3.6-27B (27B dense). Light dot: Qwen3.5-397B-A17B (397B total, 17B active).
Qwen's own harness, not an independent one — third-party replication on production coding tasks was still limited as of the release. Treat this as a vendor claim worth watching, not a settled result.

This is Qwen grading Qwen, on Qwen’s own scaffold, and independent verification on production tasks was still thin as of the release. It’s a sourced, specific result. It is not proof that dense 27-billion-parameter models generally beat 397-billion mixture-of-experts models; it’s proof that they can, on this vendor’s numbers, for this particular training run.

The dimensions a single number can’t carry

Parameter count is one axis on a grid that has gotten considerably more crowded. Context window is now a marketing number with its own asterisk: Meta advertises Llama 4 Scout at 10 million tokens, an advertised ceiling rather than a number every request reaches; Google’s Gemini 3.1 Pro model card states its own context window as “up to 1M.” Quantization changes a model’s footprint by a factor of four without changing a single weight’s meaning: a 70-billion-parameter model needs roughly 140 gigabytes of memory at 16-bit precision and closer to 35 to 40 gigabytes at 4-bit, which is the difference between needing a rack of accelerators and running on a single high-end workstation. And price per million tokens now spans five orders of magnitude on its own: Amazon’s Nova Micro lists at 3.5 cents input and 14 cents output per million tokens; Claude Fable 5 lists at $10 and $50; a legacy OpenAI reasoning model, o1-pro, priced its output at $600 per million tokens at the high end most users never touch. None of context window, quantized footprint, or price per token is derivable from parameter count alone, and all three matter more than parameter count for the decision most people are actually making, which is which model to call, not how big it is.

What “how big” was actually asking

Put the three findings next to each other and the pattern is that “how big is AI” was never one question. The physically largest disclosed model is a fact (Kimi K3, 2.8 trillion total parameters). The most consequential “biggest” models by influence and coverage are an absence of a fact (GPT-6 Astra, Claude, Gemini: no disclosed number exists for any of them). And the model an actual downloading user is most likely to pick, measured across hundreds of thousands of real choices on the largest public model repository, is under a billion parameters, by a margin of 83 to 1 percent. A single number was never going to answer all three of those, and reporting on “the biggest AI model” as if it were one story is how you end up believing something true about Kimi K3 and calling it a fact about GPT-6 Astra, when the second one was never measured at all.


References

  1. Moonshot AI. Kimi K3 model repository. github.com/MoonshotAI/Kimi-K3. Announcement and weight-release dates per Tom’s Hardware, July 2026.
  2. DeepSeek. DeepSeek-V4 announcement. Hugging Face, April 2026.
  3. Meta AI. Llama 4: Multimodal Intelligence. Meta AI blog, 2026.
  4. Google. Gemma 4 E2B model card. April 2026.
  5. Apple. Introducing the third generation of Apple foundation models. Apple Machine Learning Research, June 8, 2026.
  6. Hugging Face. State of Open Models: Summer 2026. Hugging Face blog.
  7. OpenRouter. Model rankings, checked September 15, 2026. Weekly token-volume and Series B figures per BusinessWire, May 26, 2026.
  8. Garg, A., Gaba, U., Muthirayan, D., & Chowdhury, A. R. (2025). Emissions and Performance Trade-off Between Small and Large Language Models. 2025 3rd International Conference on Foundation and Large Language Models (FLLM), IEEE. DOI: 10.1109/fllm67465.2025.11391221.
  9. Alibaba Qwen Team. Qwen3.6-27B model page. Coverage of the official release via MarkTechPost, April 22, 2026.
  10. Google DeepMind. Gemini 3.1 Pro model card.
  11. Xiong, L., Wang, H., Chen, X., Sheng, L., Xiong, Y., Liu, J., et al. (2025). DeepSeek: Paradigm Shifts and Technical Evolution in Large AI Models. IEEE/CAA Journal of Automatica Sinica, 12(5), 841-858. DOI: 10.1109/jas.2025.125495. Ten authors total; cited for confirmed DeepSeek-V2/V3 total-vs-active parameter figures as historical/background architecture data.