Mistral Large 4 Open Weights: Benchmarks and What It Needs

Mistral has put the weights of Mistral Large 4 on Hugging Face. The Paris company's flagship is a mixture-of-experts model of roughly 1 trillion parameters that activates 52 billion per token, reads text and images, and, Mistral says, is now the strongest open model for agentic coding and cyber defence.

Quick summary

• Released with open weights on October 29, 2026, three weeks after an October 6 preview.
• About 1.05 trillion parameters in total, 52B active, 1M-token context, multimodal input.
• Terminal-Bench 4.0 score rose from 28% at preview to 41%.
• A quantised version runs on four Nvidia B200 or B300 GPUs. API pricing is not yet published.

What is new in Mistral Large 4?

The Next Web reports that three extra weeks of reinforcement learning, guided by customer feedback, produced the jump. Mistral also added low and medium reasoning modes next to high, and says the high mode is 50% more efficient. Customers asked for more control over how long the model thinks, and some already use it for automated penetration testing of their own apps.

How does it score on benchmarks?

These figures come from Mistral's charts, built on Artificial Analysis indexes, so read them as vendor-supplied. AlexWrites adds that in Surge's blind tests it beat GLM-5.3, Kimi K3 and DeepSeek V4.1 Flash but won only 22% of comparisons against Claude Opus 5.5, and that Anthropic's Mythos 5.1 reaches 60.9% on Terminal-Bench 4.0 per Fortune.

Can you run Mistral Large 4 at home?

Not on a gaming PC. Even the quantised build needs four B200 or B300 data-centre GPUs. Mistral says smaller distilled versions will follow in the coming months, including one meant for laptops, but it has not committed to dates. Until then most people will reach it through Hugging Face-hosted services, Mistral's API, or Vibe.

Why does the open release matter?

Pierre Stock, Mistral's VP of Science, argued that AI will be like electricity and nobody should rely on a supplier who can cut access overnight. He pointed out that once weights are online they cannot be withdrawn, though rivals argue open models carry more risk. Mistral says it trained Large 4 on 3,800 Nvidia Grace Blackwell GPUs in Europe, and CEO Arthur Mensch notes it raised roughly $6 billion in three years versus hundreds of billions at other labs.

For context on the closed side of the race, see our GPT-6.1 Terra pricing guide, the Gemini 4 Carbon explainer and our look at DeepSeek's open-source agent.

Verdict

Large 4 is the most capable open-weight model on offer for coding agents right now, by Mistral's own numbers, but it trails the best closed models and needs serious hardware. Wait for independent tests and the API price before switching.

Related articles

DeepSeek Super-agent: Open-Source Rival to Claude CodeGPT-6.1 Terra: Price, Context and How It Compares to SolAnthropic IPO Explained: ANT Ticker, Revenue and DateGemini 4 Carbon: Access, Price and How It Compares