Mistral Large 4: 1.05T-Param Open MoE in Preview at $0.68/$2.09, Weights October 27
Mistral AI put Mistral Large 4 into public preview on October 6, 2026. The company has named it Le Chonk. The model is a granular Mixture-of-Experts with 1.05 trillion total parameters and 49 billion active during any single forward pass, alongside a 1.6-billion-parameter vision encoder that gives it native multimodal capability. Context window is one million tokens — four times the 256K limit of Mistral Large 3.
The preview API is live on Mistral Studio under the identifier mistral-large-4. Open weights are scheduled for October 27, 2026, following a staged red-teaming period.
Architecture
The granular MoE design distributes each inference pass across a large number of smaller expert blocks. At 49B active parameters per pass, per-token inference cost is comparable to dense models a fraction of the total size. The 1.6B vision encoder handles images natively, without a separate adapter or OCR preprocessing step.
Total: 1.05T parameters. Active per pass: 49B. Vision encoder: 1.6B. Supports multimodal input at launch.
Pricing
Preview pricing at the Standard tier:
| Token type | Price per million |
|---|---|
| Input | $0.68 |
| Cached input | $0.07 |
| Output | $2.09 |
At $2.09 per million output tokens, Large 4 sits substantially below closed frontier models — Claude Opus 5.5 is $20/M output, GPT-6 Astra is $20/M output — and below DeepSeek V4 Pro on hosted APIs. Mistral describes this as preview pricing; commercial rates were not confirmed at launch.
Benchmark Numbers
Launch materials report the following results, all vendor-reported and not independently verified as of October 6:
| Benchmark | Mistral Large 4 | Context |
|---|---|---|
| DeepSWE v1.1 | 62% | Kimi K3: 68%, DeepSeek V4.1 Flash: 74.2% |
| Finance Agent v2 | 54.7 | GPT-6 Astra: 53.5, Claude Opus 5.5: 58.6 |
| Finch / FinWorkBench | 67% | Finance domain reasoning |
| Harvey Legal Agent | ~15% | Legal document workflow agent |
| DIOR-RSVG | 73% | Visual grounding, spatial reasoning |
| Dense200 | 42% | Long-context dense retrieval |
Mistral’s chief scientist Guillaume Lample posted on X that Large 4 outperforms GLM 5.3 on DeepSWE and Kimi K3 on Terminal-Bench 4, without specifying the Terminal-Bench 4 score. No independent replication was available at the time of writing.
The Finance Agent v2 result is the strongest cross-lab comparison in the launch materials. Beating GPT-6 Astra by 1.2 points on a domain-specific task that Mistral targets for commercial deployment is a defensible claim. Claude Opus 5.5 leads that benchmark at 58.6.
SWE-bench Verified was not reported in launch materials.
Where It Fits in the Open-Weight Tier
Mistral is calling Large 4 the best open-weight model from the US or Europe on aggregated benchmarks — a narrower claim than frontier leadership overall, and an accurate one at current data.
The open-weight coding frontier is Chinese-dominated. Reflection Beam, which launched October 6, posts 80.9% on SWE-bench Verified. Kimi K3 holds 93.4% SWE-bench. DeepSeek V4 Pro is at 96.4% SWE-bench. On DeepSWE specifically, Mistral Large 4 at 62% trails Kimi K3 (68%) and DeepSeek V4.1 Flash (74.2%) by a meaningful gap.
DeepSeek V4 Pro uses an identical active parameter budget — 49B — at a larger total scale: 1.6T parameters against Mistral’s 1.05T. Both are granular MoE designs built for efficient frontier-quality inference.
The October 27 open weights release will allow independent evaluations on SWE-bench and tau-bench, which is where the composite score gap between Large 4 and Chinese peers will become quantifiable.
The Open Weights Argument
Mistral is framing Large 4 around AI sovereignty. The pitch is not primarily benchmark-competitive at the frontier; it is that model weights you can run and inspect cannot be turned off or guardrail-blocked by a vendor after deployment.
The concrete example Mistral referenced: when Hugging Face was targeted by rogue OpenAI agents earlier this year, it turned to American closed models for help and found their guardrails blocking defensive security work classified as attack behaviour. Hugging Face ended up using a Chinese open-weight model. Mistral’s position is that open weights from a European or American source give security teams the same operational freedom without routing sensitive workloads through Chinese infrastructure.
That argument applies selectively. It is most coherent for security, legal, and regulated-industry deployments where vendor lock-in risk is high. For general-purpose agentic coding, the benchmark gap relative to Chinese open-weight leaders is real.
Ticker Status
Mistral Large 4 is not in the Stack Futures ticker pending independent SWE-bench Verified or tau-bench results. The current DeepSWE v1.1 figure at 62% is auxiliary and insufficient for composite scoring under the current methodology. Bootstrap consideration follows the October 27 weights release and first independent evaluation.