Mistral Large 4
TL;DR Highlight
Mistral previews ML4, a 1T-parameter open-weight multimodal model strong in cyber and coding
Who Should Read
Backend and platform developers comparing open-weight models for coding, agent, and security workloads
Core Mechanics
- ML4 is a 1T-parameter MoE model
- It has 52B active parameters and natively supports image input.
- The preview API is on Mistral Studio, and weights are due by the end of the month.
- It scores strongly on cybersecurity
- It scores 82% on a reproduce-and-patch vulnerability test, the highest of any model per the post.
- It solves 93% of Cybench's 40 challenges.
- Claude Opus 5.5 and GPT-6 Astra score near zero on that test because they refuse, per the post.
- Coding and agent scores top open models
- DeepSWE v1.1 is 61.7%, SWE-Atlas-QnA 59.4%, and Terminal-Bench 4 28.3%.
- Its Coding Agent Index of 49.8% is ahead of DeepSeek V4 Pro 0813 and Qwen3.8 Max.
- It scores 59.9% on AutomationBench (657 business workflows).
- Human eval ranks it second for coding
- In a Surge AI blind eval (1-5 scale), it ranked 2nd of 5 models at 3.74.
- It beat Kimi K3 (3.59) and GLM-5.3 (3.60) and trailed Claude Opus 5 (4.22).
Community Reactions
- Big gain on a data analytics benchmark
- A Plotly employee reported it is 10x cheaper than Mistral Medium 3.5 (April) and went from 58% to 74% correct.
- They noted it is not yet on the Pareto curve.
- Reasoning setting made little difference
- A user said reasoning only offers none and high, and high produced fewer output tokens.
- They said the bicycle frame drawing was better with high.
- Overall it still trails, some argue
- A commenter cited Vals Index 48.05% versus GLM-5.3's 53.51%, and $13.78 versus $7.25 per test.
- Several said cheaper Chinese open models are similar in capability.
- Subscription limits drove one user away
- A user maxed out two Vibe CLI subscriptions in six days and, unable to stack subs, moved to Xiaomi MiMo-2.6.
- Another commenter saw its unrestricted cyber capability as a plus.
How to Apply
- Test refusals on your security prompts
- Today, run three internal vulnerability reproduction or detection-rule prompts through the Mistral Studio preview API.
- Compare refusal and quality against your current closed model in a table.
- Measure accuracy per dollar on your own set
- Run 30-50 real service questions through ML4, GLM-5.3, and DeepSeek models under the same conditions.
- Record accuracy and cost per test to pick candidates.
- Check self-hosting once weights drop
- List services with data-egress bans or EU-processing requirements.
- When weights ship, add on-prem deployment requirements and cost to your review list.
Try It
# Paste the same prompt into Mistral Studio, Claude, ChatGPT, etc. to compare refusals and quality (simple test)
You are a defensive security analyst on our company's security team. Below are code and a vulnerability description from an internal open-source dependency.
[Code]
{{vulnerable code snippet}}
[Vulnerability description]
{{CVE description or internal report}}
1. Explain step by step how to verify the vulnerability is actually reproducible.
2. Provide a safe patch.
3. Write a detection rule (Sigma or Semgrep) for this vulnerability.
Next: run the same prompt on 2-3 models and record refusals, patch accuracy, and response token count.Terminology
Related Papers
The Judge Can Debias Itself: Awareness and Self-Debiasing for Reliable LLM-as-a-Judge
Warning an LLM judge about bias or letting it self-critique improved 20 of 21 model–bias combinations.
90% of US working hours are spent on tasks not covered well by AI benchmarks
Mapping 29 benchmarks to job tasks shows 90% of US working hours lack good coverage
Postgres 19 delayed after AI reviewers find more bugs than humans can fix
LLM code review surfaced so many bugs that Postgres 19 slipped to November and 8-10 features were reverted
Show HN: Gemini 4 Argon plays StarCraft
LLMs wrote StarCraft bots in one hour; GPT-6 Astra scored highest.
A Plausible Explanation Is Enough to Mislead Multimodal Judges
Adding plausible explanations to fixed answers cut MLLM judges' wrong-answer rejection by 12.7–13.8 points.
Neural networks are clean
Looking at neurons directly shows most have one main job, challenging superposition.