Mistral Large 4 Open Weights: 41% Terminal-Bench, Half Price

The short version

On October 29, 2026, Mistral AI published the weights for Mistral Large 4 — the 1-trillion-parameter model it nicknamed "Le Chonk" — just 23 days after the public preview. The company's claim: three more weeks of reinforcement learning turned a decent open model into the best open-weight coding agent available, and the price is now half what it was in preview.

Both claims hold up, with caveats that matter if you are choosing a model today. This piece walks through the numbers, the licensing, and the hardware reality.

What actually changed in three weeks

These are Mistral's own figures, measured against its October 6 preview:

Terminal-Bench 4.0: 28.3% → 41% (+13 points), with 30% fewer output tokens
Artificial Analysis Coding Agent Index: 48%, effectively tied with the leading open model GLM-5.3 while consuming 47% fewer tokens per task
Cognition FrontierCode 2.0 (private benchmark: would a maintainer merge the agent's change?): 19.7%, the top score
BrowseComp: 84.5%, up from a 75% SFT checkpoint, ahead of GLM-5.3 and DeepSeek V4.1 Flash
FinWorkBench (Finch): 73.3%, second among open models behind Kimi K3
Artificial Analysis Cyber Index: 51% for finding, reproducing, and patching vulnerabilities without breaking existing functionality
B3 Agent Security Benchmark: 97.1% attack resistance, up from 93.3% in preview

If you read our October 12 review of the preview, you'll remember the verdict was "strong at security, mid-pack everywhere else, and the weights aren't out." The weights are out, and the coding gap closed. The general-intelligence gap did not.

Where independent numbers disagree with the launch post

Mistral reports 48% on the Coding Agent Index; Artificial Analysis independently places Mistral Large 4 at 26 on its Intelligence Index v5 in High reasoning mode. Reporting from Techstrong.ai and Fortune frames that as the highest score for any open-weight model built outside China — above Meta's Muse Spark Open, below China's DeepSeek V4.1 Flash and GLM-5.3.

Both statements are true at once. Large 4 is now competitive at agentic coding and still trails the Chinese open leaders on aggregate capability. Pick it for coding agents, not because you want the smartest open model per se.

Pricing: half the preview rate, and cheaper per task

Input: $0.70 per million tokens
Cached input: $0.07 per million tokens
Output: $2.10 per million tokens

Artificial Analysis measured operating cost per indexed task falling from $5.54 to $2.06 — a 63% drop. That attacks the specific criticism from the preview, where Large 4 cost more than four times as much per task as open models of similar intelligence.

Reasoning modes

Large 4 now exposes three settings so you can trade accuracy for spend:

Low — simple tasks; replaces the old "None" mode
Medium — everyday work
High — the default, for the hardest cases

For high-volume batch jobs, dropping to Medium (or Low where it survives evaluation) is the obvious lever. The token-efficiency gains mean the arithmetic is no longer "open model plus a huge discount," which was rarely true for trillion-parameter MoEs.

Self-hosting: possible, but not on your workstation

The weights are ungated on Hugging Face. Sizes to plan around:

FP8 (mistralai/Mistral-Large-4-1T-A52B): 68 files, ~1,060 GB
NVFP4 (-NVFP4): 32 shards, ~610 GB, Blackwell-only
Unsloth GGUF UD-IQ1_S: ~233 GB at 1-bit

Validated serving configurations are all multi-GPU. vLLM's recipe covers 8×H200 for FP8 (TP16 also works) and 4×GB300 or 8×B200 for NVFP4. SGLang lists 8×B300 (FP8) and 4×GB300 (NVFP4) as verified. For a single machine, the Hugging Face launch post notes 2-bit quants run on one Apple Silicon Ultra with 512 GB unified memory — roughly 15 tokens/sec generation on Q2_K. A 128 GB DGX Spark or any 24 GB/48 GB card cannot load even the smallest file.

The architecture explains the mismatch: 1.05T total parameters, 49B active per token (52B including embeddings and output), 72 layers, 128 routed experts with top-4 routing plus one shared expert, MLA attention with a DSA sparse indexer, a Pixtral vision encoder, and a bundled DSpark drafter for speculative decoding. Sparsity cuts compute, not memory — every expert has to be resident.

The license is the interesting part

Mistral Ecosystem Open-Weight License, an MIT derivative. You can run, modify, fine-tune, distribute, and commercialize, with no revenue cap. The single restriction: businesses operating a model-as-a-service offering that sells third parties inference or post-training access with meaningful control over inputs, parameters, or training data need a separate commercial agreement with Mistral. Internal use, embedding the model in a feature, relaying to another host, and Mistral-hosted use are carved out.

Compare that to Meta's Muse Spark Open license, which adds an attribution requirement ("Built with Muse Spark Open") and a $20M revenue trigger. Mistral's is narrower in scope and has no dollar threshold, which makes it easier to reason about for a product team — but the "selling API access" boundary deserves a legal read before you launch anything metered.

How it should change your stack

Cutting agent costs: Large 4 at $0.70/$2.10 with 47% fewer tokens per task is now the default open-weight candidate for coding agents.
Maximum accuracy: Anthropic's Claude Opus 5.5 (60.9% on Terminal-Bench 4.0) and Google's Gemini 4 Carbon (59 on the AA Index) remain ahead, and Surge Code's blind pairwise tests ranked Large 4 behind Opus 5.5 and ahead of the other open models.
Regulated deployments: this is the strongest case for Large 4. It was trained from scratch on European infrastructure with a disclosed lineage, and it is available on Microsoft Foundry — Microsoft is the first cloud provider to carry it. La Banque Postale, an existing partner, called technological sovereignty "not optional" for a public bank.
Individuals and small teams: start on the API or OpenRouter. Buying 512 GB of unified memory to self-host a 1-bit quant is a hobby decision, not an engineering one.

Bottom line

Mistral closed both gaps we flagged in our preview review — the weights are out, and the cost per task came down 63%. For agentic coding, Large 4 is now a serious first choice among open models, and the sovereign-AI pitch gives it a real enterprise wedge that has nothing to do with benchmark scores.

What it has not done is overtake the Chinese open-weight leaders on general capability, or the Anthropic and Google flagships on absolute accuracy. Read the Coding Agent Index and the Intelligence Index as separate claims and the marketing sorts itself out.