Mistral Large 4
TL;DR Highlight
Mistral Large 4: 1T-parameter multimodal model, API now, open weights end of October
Who Should Read
Backend or platform developers evaluating a model swap or self-hosting open weights
Core Mechanics
- 1T total, 49B active
- The model has 1T parameters and is natively multimodal.
- Only 49B active parameters are used per inference.
- Benchmark claims
- Mistral says it is the best open weights model from the US or Europe on aggregated benchmarks.
- The post gives no scores or comparison numbers.
- Critical workloads and visual grounding
- Mistral says it is state-of-the-art on cyber defense, manufacturing and finance.
- It says the model surpasses closed frontier models on visual grounding (locating objects in images).
- Deployable from Europe
- Mistral says it was built in Europe end-to-end.
- It can be deployed via Mistral's own Mistral Cloud infrastructure in Europe.
- Availability
- The API is available to everyone today.
Community Reactions
How to Apply
- Run your existing eval set through the API
- Send 20-50 prompts you already use to Mistral Large 4 and compare with your current model.
- Check the exact model ID on the docs.mistral.ai model page.
- Test visual grounding on your image tasks
- Run screenshot or document-image localization tasks side by side with your current model.
- The claim is versus closed models, so measure it yourself.
- Add European deployment to your shortlist if you need data residency
- If data must stay in Europe, consider Mistral Cloud as a candidate.
- For self-hosting, size GPUs for 49B active parameters once weights are released at the end of October.
Try It
# Paste into Claude or ChatGPT: build an evaluation plan for adopting Mistral Large 4
I run a service that uses {{current model}}. The main workload is {{task type, e.g. financial document summarization, image analysis}}.
I'm deciding whether to switch to Mistral Large 4 (1T parameters, 49B active, multimodal, API available, open weights at the end of October).
Create:
1. Evaluation criteria and scoring rubric for a 50-sample comparison
2. 5 visual grounding test cases, if I have image inputs
3. One table template comparing latency, cost and accuracy
Next: fill in the table, paste the results back, and ask whether to switch.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.