Mistral Large 4 Is Now Free to Download on Hugging Face

Mistral AI has released the weights of its new flagship model, Mistral Large 4, on Hugging Face — making a roughly 1-trillion-parameter multimodal system available as a free download to anyone who wants to run it on their own hardware. The model, informally nicknamed “Le Chonk,” accepts both text and image inputs and generates text, and it carries a 1-million-token context window that ranks among the largest offered by any open-weight model available today.

Despite the headline parameter count, Mistral Large 4 uses a Mixture-of-Experts (MoE) architecture that keeps inference costs manageable. Of its approximately 1 trillion total parameters, only about 49 billion are active on any given token — the model routes each token through four of 128 available expert networks, plus one shared expert, for a total of five active experts per inference pass. That active-parameter count matches DeepSeek V4 Pro, which has 1.65 trillion total parameters, meaning the two models should have roughly comparable inference costs in practice despite the large gap in total size.

What Mistral Is Emphasizing

Cybersecurity is the clearest differentiator Mistral is putting forward. Mistral Large 4 scores 82% on the Artificial Analysis Cyber Index’s vulnerability-reproduction-and-patching task, which requires a model to reproduce a real flaw in open-source software and then fix it. The model also solves 93% of the 40 challenges in Cybench, a suite of capture-the-flag-style security competition problems, and posts 93.3% attack resistance on Lakera’s B3 prompt-injection benchmark.

Mistral notes that several leading closed-source models score near zero on the vulnerability-reproduction benchmark — not because they lack capability, but because their content filters refuse to engage with tasks that resemble offensive security work, even when the goal is remediation. Open-weight models sidestep that problem entirely: the user controls both the usage policy and the deployment environment. Mistral also confirmed in internal testing that the model performs well on malware analysis, vulnerability prioritization and detection-rule writing, none of which it was explicitly trained for.

On coding, Mistral Large 4 scores 61.7% on DeepSWE v1.1, 59.4% on SWE-Atlas-QnA, and 28.3% on Terminal-Bench 4, for a combined Coding Agent Index of 49.8% — placing it ahead of DeepSeek V4 Pro 0813 and Qwen3.8 Max on that composite score. However, the coding picture is competitive: Kimi K3 scores 68.0% and DeepSeek V4.1 Flash scores 74.2% on DeepSWE v1.1, both above Large 4’s mark. On the broader Artificial Analysis Intelligence Index, Large 4 scores 38.4, trailing Kimi K3 at roughly 44 and Qwen 3.8 Max at 45.4.

Architecture and Access

Under the hood, Mistral Large 4 has 72 decoder layers, a hidden dimension of 6,144, and 64 attention heads. Its attention blocks combine Multi-head Latent Attention with a DeepSeek Sparse Attention indexer that selects up to 2,048 key positions per query — a technique also seen in models like GLM 5.3. A 1.6-billion-parameter Pixtral vision encoder handles image input using 14×14 patches. The whole architecture is designed around two types of sparsity: one over expert networks, one over context positions.

The model ships in FP8 and NVFP4 quantization variants, with day-zero inference support for Hugging Face Transformers, vLLM and llama.cpp. Community quantizations in GGUF format are available from Unsloth and the llama.cpp team. The NVFP4 variant requires NVIDIA Blackwell GPUs. A 2-bit GGUF quantization can run on a single Apple Silicon Ultra with 512 GB of unified memory — impressive, but still well beyond typical student hardware.

The license is the Mistral Ecosystem Open-Weight License, an MIT derivative. Personal use and commercial product development are free; operating the model as a third-party API service requires a separate commercial agreement with Mistral. For students who want API access without running their own hardware, the model is available through Mistral’s own platform at $0.70 per million input tokens and $2.10 per million output tokens, and also through OpenRouter and Microsoft Foundry.

Why This Matters for Students and Early-Career Researchers

The cybersecurity angle is particularly relevant for students pursuing security, software engineering or AI roles. Closed-source models from OpenAI and Anthropic frequently decline to engage with vulnerability research and incident-response tasks — the exact workflows that appear in entry-level SOC analyst and security engineering jobs. Mistral Large 4 gives students access to a frontier-class model that will actually attempt those tasks, without a per-token meter running or the risk of an API provider revoking access mid-project.

The Hugging Face post also points out a practical path for community builders: using the TRL library, developers can fine-tune smaller distilled models on traces generated by Mistral Large 4 against their own codebases. Vulnerability-reproduction tasks produce verifiable rewards, which makes them a natural fit for reinforcement-learning pipelines aimed at training smaller, more deployable security models. That pipeline — big open model generates data, smaller model trains on it — is increasingly how the research community builds specialized tools without frontier-model compute budgets.

Mistral has indicated that smaller distilled versions of Large 4 are coming, including at least one designed to run on a laptop, though no firm release dates have been announced.

For students preparing for CTF competitions, building security portfolios, or just looking for a capable coding assistant they can run without worrying about usage caps, the free weight download is the most direct on-ramp. The weights require no approval and are available now at the Mistral Large 4 collection on Hugging Face.

Source: Hugging Face