AIData brief
Ten open-weight models passed 1T parameters. All ten came from China
Epoch AI lists no open-weight model above 1T from outside China since 2025; Mistral Large 4 would break that streak once its weights ship.
Tec N Trend data desk · 1 min read · 3 open sources
- Mistral Large 4 (total params)
- 1 trillion
- Open-weight 1T+ models, all Chinese
- 10
- Largest earlier Mistral in Epoch (Pixtral Large)
- 124 billion
- Size jump vs that model
- about 8x
Mistral announced Large 4 on 6 October, a model with 1 trillion total parameters, according to TechCrunch and The Decoder. The Decoder reports 49 billion are active per token. It is live as a preview, with open weights promised for later this month.
Epoch AI's notable-models table shows how rare that size is in the open. Since July 2025 it records ten open-weight releases at 1T or more (Kimi, Qwen, DeepSeek, Ant's Ling, Xiaomi MiMo, Meituan LongCat), and every one is from a Chinese lab. Mistral's largest earlier entry is Pixtral Large at 124 billion, so Large 4 is about 8 times bigger.
The test is delivery. Until the weights are downloadable, Large 4 sits in the "announced" column. Watch for the licence terms and for whether the 49B active-parameter count holds, since that, not the 1T headline, sets running cost.
Data, method and caveats
- Epoch AI, Notable AI Models CC BY 4.0 · captured 2026-10-07
- TechCrunch, Mistral Large 4 report news report (facts only) · captured 2026-10-07
- The Decoder, Mistral Large 4 report news report (facts only) · captured 2026-10-07
- Mistral Large 4 is not in Epoch's data (latest entry 2026-09-30). Its point (1,000 billion, 2026-10-06) comes from TechCrunch and The Decoder and is flagged as "announced, weights pending".
- Epoch records total parameters. For mixture-of-experts models only a fraction is active per token, so size is not a quality or cost ranking.
- Epoch's "open weights" flag follows its own accessibility field and includes restricted and non-commercial licences. The chart shows 67 open-weight and 31 other models of 100B+ (plus Mixtral 8x7B), from 2023.
- Most closed frontier models do not disclose size, so the "no open weights" points are only those with a disclosed or estimated count; some Epoch values are rated speculative (for example Grok 4 at 3T).
- The closest US open-weight model is Thinking Machines' Inkling at 975 billion (July 2026), just under the 1T line. Epoch also lists Google's 2021 Switch Transformer (1.57T, open weights); it is a research model, outside the 2025-26 LLM wave counted in the title.
This brief as JSON or Markdown. Reuse the chart with credit to Tec N Trend and the sources above.