Mistral Large 4 (Le Chonk): an open-weight multimodal MoE model with 1M-token context

Mistral Large 4 (aka Le Chonk) is a multimodal, multilingual Mixture‑of‑Experts (MoE) model from Mistral that accepts text and images and generates text. The model has roughly 1T total parameters, 49B active parameters, and a 1,000,000‑token (1M) context window.

It is officially released in two variants: FP8 and NVFP4. Community GGUF quantizations have been produced by Unsloth and the llama.cpp team. The model is distributed under the Mistral Ecosystem Open‑Weight License (an MIT‑derivative license that specifies limitations for model‑as‑a‑service offerings).

Why this matters

The open‑weight release gives teams and researchers full control over deployment and usage policy, which is especially important for defensive security work where reproducing and patching vulnerabilities is necessary. At the same time, the model retains refusal behaviour for harmful requests: on several cyber prompts its average refusal rate is higher than any other open‑weight model reported by Mistral.

Capabilities and benchmark highlights

• Cybersecurity: on the Artificial Analysis Cyber Index task that requires reproducing a real open‑source vulnerability and patching it, Mistral Large 4 scores 82%, the highest reported on that specific test within the index. Mistral also reports strong internal results in malware analysis, vulnerability prioritization and detection‑rule generation.
• Global ranking (as of October 29, 2026): the model is the fifth best open‑weights model on the Artificial Analysis Intelligence index.
• Coding and agent benchmarks: DeepSWE v1.1 — 61.7%; SWE‑Atlas‑QnA — 59.4%; Terminal‑Bench 4 — 28.3%. Combined Coding Agent Index: 49.8% (ahead of DeepSeek V4 Pro 0813 and Qwen3.8 Max).
• Knowledge/workflow metrics: AutomationBench 59.9%; AA‑Briefcase 1,393 Elo.
• Cybench: the model solved 93% of a 40‑challenge security competition suite.
• Lakera B3: 93.3% attack resistance, indicating robustness to certain prompt‑injection attacks.

Architecture — text decoder

• Decoder depth: 72 layers; hidden dimension: 6,144; 64 attention heads.
• Attention design combines Multi‑head Latent Attention (MLA), which keeps a compressed KV representation, with a DeepSeek Sparse Attention (DSA) indexer that selects up to 2,048 key positions per query.
• Positional strategy alternates RoPE and NoPE in a 3:1 pattern (three layers use RoPE, every fourth layer uses NoPE). KV state is shared across layers.
• The first decoder layer uses a dense feed‑forward network; the remaining 71 layers are MoE blocks. A softmax router selects four of 128 experts per token, and a shared expert runs alongside them, producing five active experts per token.
• RoPE variant: Mistral employs LeRope, a YaRN‑scaled RoPE with one learned multiplier per frequency.

Vision encoder

The model uses a 1.6B Pixtral vision encoder: 30 layers, hidden dimension 2,048, 16 attention heads. Pixtral processes 14×14 patches and uses a patch‑merge projector with spatial merge size 2.

Inference support and runtimes

Day‑0 support is provided for Transformers, vLLM and llama.cpp. GGUF quants are available from community conversions.

• Transformers: the implementation exposes AutoModelForMultimodalLM and AutoProcessor to prepare text and image inputs. The model accepts reasoning_effort values low/medium/high (default high) and requires an appropriate transformers upgrade to run.
• vLLM: supports the NVFP4 checkpoint (smaller memory footprint, requires NVIDIA Blackwell GPUs). vLLM also supports the model's DSpark drafter for speculative decoding speedups.
• llama.cpp / Llama App: multiple quants are provided (IQ2_XXS, Q2_K, Q4_K_M, NVFP4, Q8_0, BF16). Some 2‑bit quants can run on a single Apple Silicon Ultra with 512 GB unified RAM; others require distributed inference across multiple GPUs.

Representative performance numbers (hardware‑dependent, reported in the release):

• Q2_K on a single M3 Ultra: prefill throughput ~90 tokens/s, generation ~15 tokens/s.
• On 4×RTX6000 Pro: NVFP4 with 32K context, prefill 130–590 tok/s, generation ~10 tok/s. IQ2_XXS at 800K context: short contexts prefill ~500 tok/s and generate ~30 tok/s, but long contexts reduce throughput significantly.

Fine‑tuning with TRL

Day‑0 TRL support is available for supervised fine‑tuning. The release demonstrates an FP8 run on 8×B300 nodes to train multilingual reasoning behaviour using TRL, transformers and peft (LoRA configuration shown in the example). Required packages include transformers, trl and peft.

Cloud endpoints and deployed performance

Mistral Large 4 is available as an Inference Endpoint with a ready recipe for 8×NVIDIA H200 GPUs, running vLLM with FP8 weights distributed across the eight GPUs using tensor and expert parallelism, and leveraging DSpark drafts. Notable figures from that deployment:

• Supports up to 131K tokens of context and up to 16 concurrent requests.
• Single request: time to first token ~0.6 s for a 1K‑token prompt; generation speed ~65 tok/s, rising to 120–138 tok/s when DSpark drafts are used.
• Long prompts: 16K‑token prompts (four concurrent) prefilled at ~4,300 tok/s total, with time to first token 8–9 s.
• Under load: about 480 tok/s total generation across 1,000 ShareGPT conversations. The endpoint is deployable with the hf CLI: hf endpoints catalog deploy --repo mistralai/Mistral-Large-4-1T-A52B --name mistral-large-4. The OpenAI‑compatible API returns generation content and a reasoning trace; the reasoning_effort parameter adjusts how much the model reasons before answering.

Dell Enterprise Hub

Dell offers optimized deployment recipes for Dell AI Servers (NVIDIA H200, B200, GB200, GB300) via the dell‑ai CLI. Example usage installs dell‑ai and runs a model deploy recipe on B300 hardware.

Where to get the model and community artifacts

All official checkpoints are available in the Mistral Large 4 collection. Community‑built GGUFs, fine‑tunings, quants and other artefacts are published or expected to be shared on the Hub.

Conclusion

Mistral Large 4 (Le Chonk) is a large, open‑weight multimodal MoE model with a very large context window and multiple supported runtime options. Its strong cybersecurity, coding and multimodal benchmark results, together with day‑one support across major inference and fine‑tuning stacks, make it a significant addition to the open‑weight ecosystem for both research and defensive security applications.