Mistral ML4 Weights: 47% Fewer Tokens Than GLM-5.3
In this article
Mistral has released weights for Mistral Large 4 (ML4), its trillion-parameter open-weight model, on Hugging Face and through the company's API. Nicknamed "Le Chonk," the model was introduced earlier in October 2026. The release lets enterprises and government agencies download, self-host, and run the system without routing through Mistral's infrastructure.
Pierre Stock, Mistral's VP of Science, told TechCrunch that an open-weight model provides "access that can't be cut off by definition." For security teams weighing local deployment, that revocation resistance is an architectural choice, not just a licensing detail: whether a local deployment is the right choice.
Scale, positioning, and the GLM-5.3 comparison
Mistral positions ML4 against GLM-5.3, which was previously considered among the strongest open-weight models for cyber work. Mistral reports ML4 achieves comparable performance to GLM-5.3 while using 47 percent fewer tokens per task. On agentic coding, Mistral claims ML4 outranks all other open models on Artificial Analysis' Coding Agent Index. Both figures are vendor-reported; TechCrunch does not cite independent replication.
CEO Arthur Mensch frames the efficiency claim as structural: Mistral says it achieves "state-of-the-art performance" at a "fraction of the compute and capital committed by closed labs." That remains a company claim, but if the 47 percent token gap holds under evaluation, it would directly affect inference cost in high-volume threat detection pipelines.
Deployment model and cybersecurity use case
Mistral targets agentic security operations — identifying, coordinating, and responding in real time to incidents — plus code and IT infrastructure analysis. Stock told TechCrunch that open sourcing means companies can deploy ML4 on-prem in their own ecosystem. That aligns with infrastructure isolation as a security primitive: where the model runs enforces some behaviour boundaries, not just its guardrails.
Stock also acknowledged dual-use risk, saying some malicious actors might try to leverage their own technology to attack, and that broad defensive capability should be strengthened quickly.
Open-weight vs. closed-model tradeoff matrix
| Dimension | ML4 (open weight) | Closed frontier model (API) |
|---|---|---|
| Access continuity | Cannot be revoked once downloaded | Subject to provider policy or outage |
| Data sovereignty | On-premises deployment supported | Data transits provider infrastructure |
| Token efficiency vs. GLM-5.3 | 47% fewer tokens per task (Mistral-reported) | Not directly comparable |
| Agentic coding rank | #1 open model, Artificial Analysis Coding Agent Index (Mistral-reported) | Closed models not ranked on same index |
| Weight availability | Hugging Face and Mistral API | Not applicable |
AI Mastery analysis
The structural fact is a trillion-parameter model now downloadable. But a capable model does not automatically produce a safe deployment. Agentic security workloads are exactly the setting where agent security failures emerge from shared surfaces rather than model logic. The orchestration layer, tool permissions, and network boundaries around ML4 determine whether it is safer than a closed alternative or just moves the trust boundary to the operator's own infrastructure.
Stock's response to rogue agent concerns — "we don't want to fearmonger people, we think it's a problem that is solvable" — is reassuring as a posture but leaves engineering specifics open. Teams evaluating ML4 should treat the 47 percent token-efficiency claim as a signal to test against their own query distributions, not a fixed budget, because it is self-reported against one named competitor.
Still, open-weight releases are now part of security procurement calculations, not just research benchmarks. When a trillion-parameter model with a claimed top agentic coding rank is downloadable, the assumption that closed models are inherently more capable needs direct re-examination.
Sources
Frequently asked questions
How many fewer tokens does Mistral Large 4 use compared with GLM-5.3?
Mistral reports that ML4 uses 47% fewer tokens per task while performing comparably to GLM-5.3. The figure is vendor-reported against that named competitor, not independently replicated in TechCrunch's article. It is a testable inference-cost signal, not a fixed production budget.
Where can I download Mistral Large 4 weights?
The weights are available on Hugging Face and via Mistral's API. The release targets enterprises and governments that want to deploy the trillion-parameter model on their own infrastructure.
Does Mistral Large 4 support on-premises deployment?
Yes. Mistral's VP of Science told TechCrunch that open sourcing means companies can deploy ML4 on-prem in their own ecosystem. An open-weight model offers access that cannot be cut off.
How does Mistral Large 4 rank on the Artificial Analysis Coding Agent Index?
Mistral claims ML4 outranks all other open models on agentic coding tasks in the index. The ranking is company-reported and appears in TechCrunch's coverage, not as independent replication.
Related Reading
Mistral Claims 82% Cyber Fix Score for 1T Le Chonk
Mistral's 1T-parameter Mistral Large 4 preview posts an 82% cyber fix score, with open weights promised by the end of October 2026.
GPT-6 Astra Hits Amazon Bedrock With 1M-Token Context and Critical Security Tier
GPT-6 Astra is now generally available on Amazon Bedrock with a 1M-token context window, chip-level operator isolation, and OpenAI's first Critical cybersecurity classification.
Z.ai GLM-5.3: Benchmark Gains From Post-Training Alone
GLM-5.3 reuses the 743B GLM-5.2 base model unchanged. Every benchmark gain comes from scaled post-training environments and longer training runs.