Mistral Open-Sources Large 4: 1T-Parameter MoE Deployable on 4 GB300s
Mistral released the weights of Mistral Large 4 (ML4, nicknamed Le Chonk) on October 29, followed by an announcement from core researcher Pierre Stock and a technical report detailing the development process. This is the official open-weights release following the Preview version three weeks earlier, and is reportedly more efficient and compact than the Preview.
Confirmed
- Model specs: A multimodal MoE model with 52B activated parameters and roughly 1T total parameters, built for reasoning and long-horizon agent tasks; the MoE architecture activates 4 of 128 routed experts per token, plus 1 shared expert (as relayed by @AIFlowML)
- Headline capabilities: Code agents and cybersecurity are described as frontier-level (per Pierre Stock)
- Deployment requirements: About 606GB when quantized, deployable on 4 GB300 GPUs
- Key points from the technical report: Quantifies the performance loss that KV-cache compression causes on downstream agent tasks (per m3/m4's summary of the report)
- Context length: About 1 million tokens (per @AIFlowML)
Why it matters
- The sparse MoE design—1T total parameters with only 52B activated—combined with quantized deployment on just 4 GB300 GPUs dramatically lowers the practical barrier to deploying ultra-large models
- The technical report's quantitative analysis of how KV-cache compression affects agent tasks offers valuable reference for engineering long-context/agent systems
- It marks another major open-weights move by Mistral, Europe's leading model lab, directly targeting the code and security capabilities of frontier closed-source models
2026-11-01 ~ 2026-11-03 · 8 related posts
- Episode 1: Mistral Open-Sources Trillion-Parameter Flagship Mistral Large 4(2026-10-29, 58 posts)
- Episode 2: Unsloth and Hugging Face Bring Le Chonk Quantized for Local Deployment(2026-10-29, 3 posts)
- Episode 3: Mistral Large 4 Scores 26 on Artificial Analysis Index, Leads Open-Weight Token Efficiency(2026-10-29, 5 posts)
- Episode 4: Mistral Large 4 update ships with open weights, ties GLM-5.3 on coding(2026-10-30, 8 posts)
- Episode 5: Mistral Open-Sources Le Chonk, Beating Kimi and DeepSeek on Coding Benchmarks(2026-10-31, 3 posts)
- Episode 6: Mistral Open-Sources Large 4: 1T-Parameter MoE Deployable on 4 GB300s(2026-11-01, 8 posts)
- Episode 7: Mistral Coding Model Benchmarks Questioned as License Terms Draw Scrutiny(2026-11-01, 4 posts)
Primary sources
- Mistral open-weights Large 4: ~1T-param MoE with 1M context, free until you host it — AIFlow_ML ·
- Mistral Large 4 tech report: 1T-param MoE, 606GB quantized, runs on four GB300 GPUs — _AndrewZhao ·
- Mistral releases open-weight ML4 with frontier coding-agent and cyber capabilities, plus tech report — _AndrewZhao ·
- [source] Mistral open-weights Large 4: ~1T-param MoE with 1M context, free until you host it — AIFlow_ML · 2026-11-01
- Mistral Large 4 goes open-weight with frontier code-agent and cyber capabilities — _AndrewZhao · 2026-11-03
- [source] Mistral Large 4 tech report: 1T-param MoE, 606GB quantized, runs on four GB300 GPUs — _AndrewZhao · 2026-11-03
- Mistral Large 4 goes open-weight: 1T-param MoE with 52B active, runs on four GB300 GPUs — _AndrewZhao · 2026-11-03
- [source] Mistral releases open-weight ML4 with frontier coding-agent and cyber capabilities, plus tech report — _AndrewZhao · 2026-11-03
- Mistral publicly releases Large 4, a downloadable model targeting cybersecurity users — _AndrewZhao · 2026-11-03
2 near-duplicate retellings: aronchick · _AndrewZhao