Mistral ML4 open-weight release explained: Le Chonk, coding, security and reasoning costs

ML4 is short for Mistral Large 4, nicknamed Le Chonk. It did not first appear this week: the model documentation dates its public preview to October 6, 2026. The latest announcement releases weights improved through additional reinforcement learning, with updated coding, defensive-security and reasoning-efficiency results.

This article concerns that improved open-weight release, keeping preview results separate from current results. For organizations, the questions include not just capability but control, customization and the cost of running it.

Sources were checked on October 30, 2026. Scores below are reported or cited by Mistral, not independently reproduced by TboxIO. Cost examples are calculations using published rates.

Specifications: total and active parameters are different

Item	Official information
Name	Mistral Large 4 / ML4 / Le Chonk
Architecture	Natively multimodal, granular MoE
Total parameters	Documentation lists 1.05T; commonly abbreviated approximately 1T
Active parameters	Approximately 49B per token, or 52B including embeddings and output layers
Experts	128 routed experts, four active per token, plus one shared expert
Vision encoder	Documentation lists 1.6B parameters
Modalities	Text and image input; text output
Context	1M
Reasoning effort	low, medium, high; high by default
API identifier	mistral-large-4
Weight repository	mistralai/Mistral-Large-4-1T-A52B

The 49B and 52B figures use different counting scopes, as explained in the official model card. They should not be presented as two conflicting models. Activating only part of an MoE model also does not mean a deployment stores only 52B parameters.

The card claims native fluency in more than 160 languages, not identical quality in every language. Text/image input likewise does not establish native audio or video output; distinguish model modalities from tools and interfaces.

Reported benchmarks: coding efficiency, with broader limits

Metric	Latest reported result	Interpretation
Cognition FrontierCode 2.0	19.7%	Private benchmark examining end-to-end code quality and whether a maintainer would merge the change
Artificial Analysis Coding Agent Index	48%	Mistral describes an effective tie with GLM-5.3, using 47% fewer tokens per task on average
Terminal-Bench 4.0	41%	13 percentage points above public preview, with 30% fewer output tokens
Artificial Analysis Cyber Index	51%	Finding, reproducing and validating vulnerabilities, then patching without breaking functionality
FinWorkBench (Finch)	73.3%	Finance/accounting workflows, particularly spreadsheet creation and editing
BrowseComp	84.5%	Compared in the announcement with a 75% SFT checkpoint, not every prior release
B3 Agent Security attack resistance	97.1%	Preview scored 93.3%; a 3.8-percentage-point increase
Artificial Analysis Intelligence Index v5	26	An aggregate index, not 26% accuracy

These metrics have different units and setups, so averaging them into an overall win rate would be misleading. FrontierCode 2.0 is private; its score alone is insufficient for public reproduction. The 47% token reduction is an average against the specified comparison model, not a universal saving across tasks and alternatives.

The official card also acknowledges that leading Chinese open-weight models score higher on Intelligence Index v5. Strong coding and security results do not establish universal leadership.

Mistral cites Surge Code blind preferences favoring ML4 over GLM-5.3, Kimi K3 and DeepSeek V4.1 Flash, behind only Claude Opus 5.5 among the models evaluated. This is human preference within a particular evaluation set, not an absolute ranking across all coding work. We do not add unverified numerical win rates.

What is directly compared with the preview?

Terminal-Bench 4.0 is the clearest same-series comparison: 41%, up 13 percentage points, alongside 30% fewer output tokens. Subtracting the reported difference implies a preview score of approximately 28%; that is arithmetic, not a TboxIO measurement. A 13-point increase should not be confused with a 13% relative improvement.

B3 resistance rises from 93.3% to 97.1%, a 3.8-point difference. BrowseComp's rise from 75% to 84.5% explicitly compares an SFT checkpoint with subsequent reinforcement learning, not necessarily the old production API.

For other tasks, the cited text does not supply a complete controlled preview comparison. We therefore do not invent improvement multiples.

Defensive security does not mean error-free security

Cyber Index tests vulnerability discovery, reproduction, validation and repairs that preserve functionality. Mistral separately reports a 96.6% average refusal rate on malicious cyber prompts from JailbreakBench, StrongREJECT and AgentHarm.

That 96.6% refusal rate and B3's 97.1% resistance rate measure different things. Neither means discovering 97.1% of real vulnerabilities or complete immunity to prompt injection.

The announcement also describes a specialized security variant for vetted enterprise partners, with separate access requirements. It should not be conflated with the publicly downloadable general ML4. Real review and patching still require checks for false positives, missed problems and regressions; model scores do not replace permissions or verification in the surrounding system.

Three reasoning levels, high by default

Level	Official recommendation
low	Simple tasks; replaces the previous None mode
medium	Everyday tasks
high	Default; complex agents and high-precision work

The card recommends temperature=1.0 and top_p=0.95, with clearly defined, focused tools. Existing workflows should not assume the previous None setting retains the same meaning.

The model supports function calling, JSON output and strict JSON Schema. More reasoning can change latency and token use; choose based on completion quality and cost rather than assuming high is best everywhere. The card specifies that max_completion_tokens includes both reasoning tokens and visible output.

API rates and an example

The latest announcement and documentation list these USD prices per million tokens:

Item	Price
Uncached input	$0.70
Cached input	$0.07
Output	$2.10

With 10,000 uncached input tokens and 2,000 billable output tokens, model-token cost is approximately $0.0112. This fixes token counts for illustration, excludes other tools/platform charges and does not promise every task uses that amount.

Cached rates require actual cache hits. Reasoning, tool results and retries alter total usage. Reported token efficiency is relevant but does not guarantee an identical percentage bill reduction for every project. Other hosting providers may have different prices and service conditions.

Self-hosting: open weights do not imply a consumer single-GPU deployment

The official card recommends vLLM and lists SGLang, Transformers and llama.cpp ecosystem support. Its examples use eight-way tensor parallelism, which is a configuration example, not a universal minimum GPU count across precisions and hardware.

The base repository is tagged FP8, and an NVFP4 checkpoint targets Blackwell GPUs. Quantization can reduce memory, but total weights, KV cache, concurrency and context still determine deployment capacity. TboxIO has not measured local hardware speed or VRAM use.

An embedded DSpark speculative-decoding draft head is included. The card reports up to 1.9 times throughput, explicitly dependent on hardware and workload. Throughput is not per-user latency, nor does it guarantee 1.9 times speed on any device.

Tool integration needs compatible reasoning and tool-call parsers. The official vLLM example uses mistral parsers; lowering maximum context or concurrency is recommended when memory is insufficient. Supporting 1M does not mean arbitrary hardware can run a fully occupied context easily.

License and evaluation priorities

Mistral describes the Mistral Ecosystem Open-Weight License as a modified MIT license. Its announcement and model card specifically say commercial operation of a Model-as-a-Service business requires a separate Mistral commercial license. It should not be labeled standard MIT without additional conditions or Apache 2.0; read the repository's full LICENSE for actual use.

Teams needing controllable coding agents, defensive-security assistance or document/chart understanding have reason to evaluate ML4. Existing preview users should compare completion rates, human corrections, tokens, latency and costs on fixed repositories and tasks. Self-hosting teams must also assess hardware, quantization and operations rather than API rates alone.

TboxIO's assessment: the release combines improved open weights with coding, defensive-security and token-efficiency gains. Reported results justify evaluation, not claims of superiority on every task, perfect security or cheap single-GPU deployment.

Sources

• Mistral open-weight announcement: results, preview comparisons, pricing and license
• Official model documentation: parameters, context, API and pricing
• Official model card: architecture, parameter counting, reasoning and deployment