Skip to content
Theorem Weekly TW markTWTheoremWeekly

The evidence behind the week in machine learning.

Model updates

Does Mistral Large 4 Open Weights Match Closed Models in Agentic Coding?

Mistral AI releases open weights for its trillion-parameter model, claiming top-tier coding agent performance with significantly lower token costs.

Key points

  • Mistral Large 4 has 1 trillion parameters with 52 billion active parameters and is natively multimodal.
  • The model ties GLM-5.3 on Terminal-Bench 4.0 agentic coding tests with a score of 41%.
  • Training utilized 3,800 Nvidia Grace Blackwell GPUs in European data centers.
Does Mistral Large 4 Open Weights Match Closed Models in Agentic Coding?
Depth
1 trillion parametersMistral Large 4 has 1 trillion parameters with 52 billion active parameters and is natively multimodal · reported
52 billionMistral Large 4 has 1 trillion parameters with 52 billion active parameters and is natively multimodal · reported
41%The model ties GLM-5.3 on Terminal-Bench 4.0 agentic coding tests with a score of 41% · reported
1 outlet · newest source 428h ago

Mistral AI SAS announced the availability of open weights for Mistral Large 4, its most capable large language model to date, according to SiliconANGLE AI. This release follows a preview version launched three weeks prior, allowing researchers and enterprises to deploy the model on private infrastructure.

Evidence: reportedMistral releases open weights for Mistral Large 4, providing sovereign control for the company’s ‘Le Chonk’ AI model

aifleet-siliconangle

What was released or claimed

The company states that Mistral Large 4, nicknamed “Le Chonk,” features 1 trillion parameters with 52 billion active parameters. SiliconANGLE AI reports that the model is natively multimodal and underwent three additional weeks of reinforcement training and customer testing after its initial preview. Mistral claims the open-weight version delivers substantial gains with no signs of saturation in performance.

Evidence: reportedMistral releases open weights for Mistral Large 4, providing sovereign control for the company’s ‘Le Chonk’ AI model

aifleet-siliconangle

Evidence: reportedMistral releases open weights for Mistral Large 4, providing sovereign control for the company’s ‘Le Chonk’ AI model

aifleet-siliconangle

Who it affects

Organizations seeking sovereign control over AI workloads can now deploy the model within their own firewalls. SiliconANGLE AI notes this provides a Western-developed alternative to large Chinese coding models like the GLM and Kimi families. The release targets companies needing to process sensitive data without exposing it to external cloud providers or third-party services.

What is still unknown

The source material does not detail the specific reinforcement learning techniques or datasets used during the additional training period. It is unclear how the model performs on non-coding tasks compared to other frontier models. SiliconANGLE AI does not report independent verification of the company’s claims regarding cyber defense capabilities or token efficiency metrics.

So what

Mistral asserts that open weights prevent state-of-the-art AI capabilities from concentrating in the hands of a few actors. Co-founder Arthur Mensch stated that Mistral raised roughly $6 billion over three years, contrasting this with other labs that raised hundreds of billions. The company emphasizes end-to-end control of its infrastructure enabled optimization for hardware during model development.

Evidence: reportedMistral releases open weights for Mistral Large 4, providing sovereign control for the company’s ‘Le Chonk’ AI model

aifleet-siliconangle

What the paper shows

Mistral Large 4 was trained on 3,800 Nvidia Grace Blackwell graphics processing units in European data centers. On the Artificial Analysis Coding Agent Index, the model rivals GLM-5.3 while using 47% fewer tokens per task on average. On Terminal-Bench 4.0, which tests agentic coding in the terminal, ML4 improved from 28% to 41%, tying GLM-5.3 as the best model for coding agents according to the reported results.

Evidence: reportedMistral releases open weights for Mistral Large 4, providing sovereign control for the company’s ‘Le Chonk’ AI model

aifleet-siliconangle

Evidence: reportedMistral releases open weights for Mistral Large 4, providing sovereign control for the company’s ‘Le Chonk’ AI model

aifleet-siliconangle

Evidence: reportedMistral releases open weights for Mistral Large 4, providing sovereign control for the company’s ‘Le Chonk’ AI model

aifleet-siliconangle

Why it matters

  • Researchers can evaluate the efficiency gains of sparse mixture-of-experts architectures with 52 billion active parameters against dense models.
  • The reported 47% reduction in token usage on coding tasks offers a concrete metric for cost-benefit analysis in agentic workflows.
  • Deployment on private infrastructure allows for security audits of a Western-developed model competing with Chinese counterparts.
  • The improvement from 28% to 41% on Terminal-Bench 4.0 demonstrates the impact of additional reinforcement learning on agentic coding capabilities.

Topics: Mistral AI, open weights, agentic coding, LLM benchmarks, Nvidia

Updated

Ask the article

Local keyword search of this article; no generated answers.

Did this article give you the evidence you needed?

No vote saved on this device.
Your choice stays on this device. No public tally.

Questions readers ask

How many parameters does Mistral Large 4 have?

The model has 1 trillion total parameters with 52 billion active parameters.

What benchmark did Mistral Large 4 tie on?

It tied GLM-5.3 on Terminal-Bench 4.0 with a score of 41%.

Where was the model trained?

It was trained on 3,800 Nvidia Grace Blackwell GPUs in European data centers.

Connected stories

This storyCan Mistral Large 4 Reproduce Vulnerabilities Where Closed Models Refuse?Can Mistral Large 4 Reproduce VuDoes Jensen Huang’s resignation reshape Nvidia’s China compliance posture?Does Jensen Huang’s resignation Realtime-Bench evaluates AI systems on dynamic tasks where delayed actions affect successRealtime-Bench evaluates AI systNvidia interim CEO sets Nov 18 meeting amid difficult periodNvidia interim CEO sets Nov 18 m

Reading receipt

  • Mistral Large 4 has 1 trillion parameters with 52 billion active parameters and is natively multimodal.
  • The model ties GLM-5.3 on Terminal-Bench 4.0 agentic coding tests with a score of 41%.
  • Training utilized 3,800 Nvidia Grace Blackwell GPUs in European data centers.
Time spent: 0:00 · Depth: Full

Sources

  1. SiliconANGLE AI

Related stories

Background reading