Mistral Drops Large 4, a 1 Trillion Parameter Open Coding Agent

Mistral has released the full weights for Mistral Large 4, nicknamed Le Chonk, following a public preview earlier in the month. The natively multimodal mixture-of-experts model contains 1 trillion parameters, activates 49 billion per token, and focuses on coding agents and defensive cybersecurity.

Developers can run, inspect, and fine-tune the checkpoint under Mistral’s modified MIT license. The combination of open weights, a 1 million-token context window, configurable reasoning, and lower reported token usage gives teams another option for deploying agents on private infrastructure.

One trillion parameters, 49 billion at work

A mixture-of-experts model routes each token through a small subset of specialized parameter groups, reducing the computation required for inference. Mistral Large 4 uses 128 routed experts, with four selected for each token, plus one shared expert. It activates 49 billion parameters per token, or 52 billion when embeddings and output layers are included.

• Training: Built from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral’s European data centers
• Context: Up to 1 million tokens, subject to serving memory and runtime limits
• Vision: A Pixtral-style encoder that accepts images at variable resolutions
• Languages: More than 160, including every official EU language
• Workloads: Instruction following, reasoning, coding, tool use, and multi-step agents

The API exposes three reasoning levels through the per-request reasoning_effort parameter:

• Low: Fast chat responses; replaces the preview’s None mode
• Medium: General day-to-day workloads
• High: Complex agent runs and precision-sensitive tasks; enabled by default

Code benchmarks favor token efficiency

Mistral’s published results place Large 4 near the top of several coding and security evaluations, although benchmark harnesses, agent scaffolding, and private test sets limit direct comparisons across vendors.

Cognition’s FrontierCode 2.0 assesses complete repository changes and asks whether a maintainer would merge the result. On that private benchmark, Large 4 scored 19.7%, ahead of the listed results for GLM-5.3 and Kimi K3.

Surge Code’s blind pairwise evaluations also preferred Large 4 over GLM-5.3, Kimi K3, and DeepSeek V4.1 Flash. Claude Opus 5.5 ranked higher among the evaluated models, leaving a measurable quality gap to the leading closed model.

Actual cost savings will vary with prompt length, generated output, caching, serving hardware, and agent design. The reported 47% reduction in tokens per task provides a useful signal for high-volume workloads, but it does not translate directly into a 47% reduction in total infrastructure cost.

DSpark accelerates token generation

The release includes a speculative-decoding component called DSpark alongside the main checkpoint. Its three dense multi-head latent attention draft layers propose blocks of three tokens through a recurrent Markov head, allowing the target model to verify several candidates at once.

Mistral reports up to 1.9 times higher throughput than serving the target model alone. Speculative decoding preserves accepted target-model outputs, while realized gains depend on batch size, hardware, runtime configuration, and draft-token acceptance rates.

Guardrails focus on defensive security

Mistral trained the final checkpoint to distinguish defensive security work from malicious offensive requests. The company reports a 96.6% average refusal rate on its malicious cyber evaluation, the highest result among the open models in its comparison, while retaining performance on defensive benchmarks.

The cyber evaluation covers practical work such as finding vulnerabilities in a repository, reproducing them, validating their impact, and producing patches without breaking existing behavior. Mistral tested the model with CrowdStrike, Cloudflare, Rubrik, Zscaler, Fortinet, and Orange Cyberdefense on enterprise security workloads.

A specialized penetration-testing variant remains limited to vetted enterprise design partners. Because the standard checkpoint is downloadable and fine-tunable, its published refusal rates describe the released configuration; downstream training, prompts, and serving policies can change that behavior.

The tradeoffs remain visible

Large 4’s strongest results cluster around coding, terminal use, and cybersecurity. Its score of 26 on the Artificial Analysis Intelligence Index v5 leads the evaluated Western open-weight models, while several Chinese open-weight models score higher on the broader aggregate test.

Operating a trillion-parameter checkpoint also requires substantial aggregate GPU memory, storage, and interconnect bandwidth. Tensor parallelism distributes the model across devices, but the required hardware still depends on checkpoint precision, context length, cache size, and concurrency.

Serving the model

The weights are available on Hugging Face, and Mistral offers a hosted API with OpenAI-compatible request fields.

Mistral recommends vLLM for local serving and provides the following container configuration for an eight-way tensor-parallel deployment:

docker run --runtime nvidia --gpus all \
  -v ~/.cache/huggingface:/root/.cache/huggingface \
  -p 8000:8000 --ipc=host \
  vllm/vllm-openai:mistral-large-4 \
  --model mistralai/Mistral-Large-4-1T-A52B \
  --tensor-parallel-size 8 \
  --tool-call-parser mistral \
  --reasoning-parser mistral \
  --enable-auto-tool-choice

SGLang, llama.cpp, and Transformers are also supported. An NVFP4 checkpoint reduces memory use and targets NVIDIA Blackwell GPUs through vLLM and SGLang; Mistral says the quantized release retains the model’s benchmark performance. Hardware capacity should be checked against the selected checkpoint and intended context length before deployment.

Modified MIT adds a service restriction

The license permits use, modification, fine-tuning, and commercialization, including enterprise deployment on private infrastructure. Businesses that offer the model through a Model-as-a-Service product must obtain a separate commercial license from Mistral.

The added service restriction makes “open-weight” the precise description for the release. Teams planning redistribution or hosted model access should review the published terms against their product architecture.

The production case rests on control and cost

Mistral Large 4 gives engineering teams control over data location, serving infrastructure, fine-tuning, and security policy while delivering competitive results on repository-level coding tasks. Its reported token efficiency could lower operating costs for agents that generate large volumes of code or terminal output.

Deployment decisions will still depend on workload-specific tests. A useful evaluation should measure patch acceptance, regression rates, tool-call reliability, latency, token consumption, and infrastructure cost on the team’s own repositories. The published results make Large 4 a credible candidate for that evaluation, especially where private deployment and defensive security capabilities carry operational weight.