We're releasing the weights of Mistral Large 4

Oct 29, 2026•kuberwastaken•View Original

TL;DR Highlight

Mistral released Large 4 weights, claiming top open-weight scores on coding benchmarks

Who Should Read

A backend developer in a regulated industry looking for a self-hostable, non-Chinese open-weight model for coding agents

Core Mechanics

1T multimodal weights released
- Mistral Large 4 (Le Chonk) is a 1-trillion-parameter, natively multimodal model.
- Weights shipped after three more weeks of reinforcement learning (RL) following the preview.
19.7% tops FrontierCode 2.0
- Cognition's private benchmark checks whether a repo maintainer would merge the agent's change.
- Compared against GLM-5.3 Flash, Kimi K3, GLM-5.3 and DeepSeek V4.1 Flash, among others.
Same score, 47% fewer tokens
- It scores 48% on the Artificial Analysis Coding Agent Index, effectively tied with the top open model GLM-5.3.
- It uses 47% fewer tokens per task than GLM-5.3 on average.
Terminal-Bench 4.0 hits 41%, +13pp vs preview
- Mistral says this leads open-weight models.
- Output tokens dropped 30% versus the preview.
Three reasoning modes
- Low is for simple tasks and replaces the old 'None' mode.
- Medium is for everyday tasks, and High is the default for the hardest ones.
Security benchmark numbers
- Artificial Analysis Cyber Index is 51%, covering vulnerability discovery, reproduction and patching.
- Average refusal of malicious cyber requests is 96.6%, and B3 Agent Security attack resistance is 97.1% (preview 93.3%).
BrowseComp 84.5%, Finch 73.3%
- RL lifted BrowseComp from 75% at the SFT checkpoint to 84.5%.
- FinWorkBench (Finch) is 73.3%, second among open models behind Kimi K3.

Community Reactions

Price reportedly cut nearly in half vs preview
- A commenter cites $0.7/M input and $2.1/M output, versus $1.36/$4.18 for the preview.
- They say the preview wasn't competitive in their evals at the old price, but this warrants reassessment.
Pushback: it's weights-available, not open weight
- Commenters note the modified MIT license requires a commercial license for Model-as-a-Service businesses.
- They point out it isn't OSI-approved.
Significant for regulated industries that can't use Chinese models
- A commenter says US regulated businesses mostly had gpt-oss or Gemma for self-hosting.
- A near-frontier model opens more options, they argue.
Skeptical hands-on experiences
- One user's code documentation task saw it ignore SKILL.md files and hallucinate.
- Another found it more expensive and less capable than Reflection's Beam 1.0 on botsbench.com cyber tests.
- A critic notes no stat on refusing legitimate requests alongside the 96.6% refusal figure.
Hardware reality check
- A user with 4x H200 (540GB) says they'd need 3-4 bit quantization to serve it.

How to Apply

Compare costs on today's pricing
- Pick 10 of your coding-agent prompts and run them on Mistral Large 4 and your current model.
- Record accuracy and tokens per task for both.
Route by reasoning mode
- Send simple tasks like lint fixes to Low.
- Use High only for hard refactors and compare the cost.
Check license and hardware before self-hosting
- Confirm with legal whether the Model-as-a-Service license condition applies to you.
- Size GPU memory assuming quantized serving for a 1T model.
Measure both refusals and pass rate for security agents
- Build separate sets of legitimate defensive prompts and malicious requests.
- Count how often legitimate requests get refused.

Try It

copy & run

# Paste into Claude or ChatGPT to draft a model-swap evaluation plan (simplified version of the article's idea)

We are evaluating whether to switch our coding-agent LLM to Mistral Large 4 (open weights, Low/Medium/High reasoning modes).
Current model: {{current model}}
Main tasks: {{e.g., bug fixes, writing tests}}

Create an evaluation plan with:
1. Criteria for choosing 10 representative tasks
2. Per-task metrics (success, tokens, cost, latency)
3. A comparison table template across Low/Medium/High
4. Three test cases to check refusals of legitimate requests

Next step: fill the table with real results and ask for routing rules.

Terminology

Open weightThe trained numbers (weights) can be downloaded. It differs from open source, and this model carries license restrictions.

MoE (Mixture of Experts)An architecture that activates only some expert sub-networks per input. Like calling only the relevant specialist at a hospital.

RL (reinforcement learning)Training that tunes a model using outcome scores as rewards. Unlike SFT, which shows correct answers, it learns by trial and error.

Sovereign AIAn organization runs and controls the model on its own infrastructure. It matters when data-residency rules apply.

Terminal-BenchA benchmark measuring how well agents handle coding and system tasks in a terminal.

Related Resources

We're releasing the weights of Mistral Large 4

Oct 29, 2026•kuberwastaken•View Original

TL;DR Highlight

Mistral released Large 4 weights, claiming top open-weight scores on coding benchmarks

Who Should Read

A backend developer in a regulated industry looking for a self-hostable, non-Chinese open-weight model for coding agents

Core Mechanics

1T multimodal weights released
- Mistral Large 4 (Le Chonk) is a 1-trillion-parameter, natively multimodal model.
- Weights shipped after three more weeks of reinforcement learning (RL) following the preview.
19.7% tops FrontierCode 2.0
- Cognition's private benchmark checks whether a repo maintainer would merge the agent's change.
- Compared against GLM-5.3 Flash, Kimi K3, GLM-5.3 and DeepSeek V4.1 Flash, among others.
Same score, 47% fewer tokens
- It scores 48% on the Artificial Analysis Coding Agent Index, effectively tied with the top open model GLM-5.3.
- It uses 47% fewer tokens per task than GLM-5.3 on average.
Terminal-Bench 4.0 hits 41%, +13pp vs preview
- Mistral says this leads open-weight models.
- Output tokens dropped 30% versus the preview.
Three reasoning modes
- Low is for simple tasks and replaces the old 'None' mode.
- Medium is for everyday tasks, and High is the default for the hardest ones.
Security benchmark numbers
- Artificial Analysis Cyber Index is 51%, covering vulnerability discovery, reproduction and patching.
- Average refusal of malicious cyber requests is 96.6%, and B3 Agent Security attack resistance is 97.1% (preview 93.3%).
BrowseComp 84.5%, Finch 73.3%
- RL lifted BrowseComp from 75% at the SFT checkpoint to 84.5%.
- FinWorkBench (Finch) is 73.3%, second among open models behind Kimi K3.

Community Reactions

Price reportedly cut nearly in half vs preview
- A commenter cites $0.7/M input and $2.1/M output, versus $1.36/$4.18 for the preview.
- They say the preview wasn't competitive in their evals at the old price, but this warrants reassessment.
Pushback: it's weights-available, not open weight
- Commenters note the modified MIT license requires a commercial license for Model-as-a-Service businesses.
- They point out it isn't OSI-approved.
Significant for regulated industries that can't use Chinese models
- A commenter says US regulated businesses mostly had gpt-oss or Gemma for self-hosting.
- A near-frontier model opens more options, they argue.
Skeptical hands-on experiences
- One user's code documentation task saw it ignore SKILL.md files and hallucinate.
- Another found it more expensive and less capable than Reflection's Beam 1.0 on botsbench.com cyber tests.
- A critic notes no stat on refusing legitimate requests alongside the 96.6% refusal figure.
Hardware reality check
- A user with 4x H200 (540GB) says they'd need 3-4 bit quantization to serve it.

How to Apply

Compare costs on today's pricing
- Pick 10 of your coding-agent prompts and run them on Mistral Large 4 and your current model.
- Record accuracy and tokens per task for both.
Route by reasoning mode
- Send simple tasks like lint fixes to Low.
- Use High only for hard refactors and compare the cost.
Check license and hardware before self-hosting
- Confirm with legal whether the Model-as-a-Service license condition applies to you.
- Size GPU memory assuming quantized serving for a 1T model.
Measure both refusals and pass rate for security agents
- Build separate sets of legitimate defensive prompts and malicious requests.
- Count how often legitimate requests get refused.

Try It

copy & run

# Paste into Claude or ChatGPT to draft a model-swap evaluation plan (simplified version of the article's idea)

We are evaluating whether to switch our coding-agent LLM to Mistral Large 4 (open weights, Low/Medium/High reasoning modes).
Current model: {{current model}}
Main tasks: {{e.g., bug fixes, writing tests}}

Create an evaluation plan with:
1. Criteria for choosing 10 representative tasks
2. Per-task metrics (success, tokens, cost, latency)
3. A comparison table template across Low/Medium/High
4. Three test cases to check refusals of legitimate requests

Next step: fill the table with real results and ask for routing rules.

Terminology

Open weightThe trained numbers (weights) can be downloaded. It differs from open source, and this model carries license restrictions.

MoE (Mixture of Experts)An architecture that activates only some expert sub-networks per input. Like calling only the relevant specialist at a hospital.

RL (reinforcement learning)Training that tunes a model using outcome scores as rewards. Unlike SFT, which shows correct answers, it learns by trial and error.

Sovereign AIAn organization runs and controls the model on its own infrastructure. It matters when data-residency rules apply.

Terminal-BenchA benchmark measuring how well agents handle coding and system tasks in a terminal.

We're releasing the weights of Mistral Large 4

TL;DR Highlight

Who Should Read

Core Mechanics

Community Reactions

How to Apply

Try It

Terminology

Related Papers

Related Resources

We're releasing the weights of Mistral Large 4

TL;DR Highlight

Who Should Read

Core Mechanics

Community Reactions

How to Apply

Try It

Terminology

Related Papers

Related Resources