Mistral Large 4 Open Weight: Run a 1T Model Yourself

Mistral released the weights for Mistral Large 4, a 1.05-trillion-parameter Mixture-of-Experts model with roughly 49 billion active parameters per token, on Hugging Face on October 29th under the identifier mistralai/Mistral-Large-4-1T-A52B. The model supports a one-million-token context window and a Pixtral vision encoder, and vLLM shipped day-0 support the same day via the vllm/vllm-openai:mistral-large-4 Docker image with mistral-common 1.13, including an NVFP4 checkpoint for NVIDIA Blackwell GPUs plus day-0 support for AMD MI350 Series and Google TPU v7x. Mistral positioned Large 4 for cybersecurity work, where it scores 82% on the Artificial Analysis Cyber Index and solves 93% of Cybench's 40 challenges, though self-hosting requires at least 256 GB of RAM for the 233 GB 1-bit Unsloth GGUF quantization that retains 91% of top-1 accuracy, and the license is not Apache 2.0.

Mistral dropped the weights for Large 4 on October 29th — a one-trillion-parameter model, downloadable from Hugging Face, no waitlist required. That sentence deserves a moment. The largest open-weight frontier model ever released is now sitting in a public repository. The catch is that “downloadable” has real hardware prerequisites, and the license is not the permissive Apache 2.0 developers might expect. Both details matter before you spin up a deployment. What You’re Actually Getting Mistral Large 4 is a Mixture-of-Experts model with 1.05 trillion total parameters and approximately 49 billion active parameters per token. The granular MoE architecture runs 128 routed experts plus one shared expert, selecting the top four per token — which is what keeps inference computationally tractable despite the trillion-parameter headline. It supports a one-million-token context window and includes a Pixtral vision encoder for native multimodal input: images, charts, technical drawings, and documents. The official weights are on Hugging Face https://huggingface.co/mistralai/Mistral-Large-4-1T-A52B as mistralai/Mistral-Large-4-1T-A52B . Unsloth has published Dynamic GGUF quantizations ranging from a 233 GB 1-bit version to a 1.1 TB Q8 near-full-precision checkpoint. The 1-bit Dynamic GGUF retains 91% of top-1 accuracy — legitimate for production, not just experimentation. The Hardware Reality Here is where the “you can actually run it” part requires qualification. The smallest quantization needs at least 256 GB of total RAM or unified memory. That puts it within reach of a Mac Studio Ultra with a 512 GB configuration, a workstation with dual DGX Sparks, or a well-equipped cloud instance — but not a developer laptop. The 2-bit quantization needs 384–512 GB; the Q8 checkpoint needs 1.2 TB or more. | Quantization | File Size | Min RAM Required | Top-1 Accuracy Retained | |---|---|---|---| | UD-IQ1 S 1-bit | ~233 GB | 256 GB | 91% | | UD-IQ2 M 2-bit | ~302–378 GB | 384–512 GB | 95% | | Q8 near full | ~1.1 TB | 1.2 TB+ | ~99% | For most developers, self-hosting Mistral Large 4 means provisioning a cloud instance with sufficient GPU memory, not plugging in a consumer GPU. That is a meaningful distinction from models like Llama 3.1 8B or Mistral Small 4, which run comfortably on a single RTX 4090. The scale is genuinely different here. Production Deployment: vLLM Day-0 Support The practical deployment path is vLLM, which shipped day-0 support https://vllm.ai/blog/2026-10-29-mistral-large-4 on October 29th alongside the weight release. The Docker image vllm/vllm-openai:mistral-large-4 includes mistral-common 1.13 and supports the NVFP4 checkpoint on NVIDIA Blackwell GPUs for a smaller memory footprint and higher throughput. AMD MI350 Series and Google TPU v7x also have day-0 support. For local experimentation, Unsloth’s guide covers llama.cpp-based serving https://unsloth.ai/docs/models/mistral-large-4 with granular GGUF selection. Where It Actually Stands Out: Cybersecurity Mistral positioned Large 4 explicitly for cybersecurity work, and the benchmarks back that up. On the Artificial Analysis Cyber Index — a test that asks a model to reproduce a real vulnerability in open-source software and then patch it — Large 4 scores 82%. On Cybench, a set of 40 capture-the-flag-style security challenges, it solves 93%. Several closed frontier models decline to attempt the vulnerability reproduction test. An open-weight model that outperforms on security tasks while remaining self-hostable is a combination that did not exist before this release. In Mistral’s internal testing, the model demonstrated capability in malware analysis, vulnerability prioritization, and detection rule authoring — tasks it was not explicitly trained for. That generalization, combined with the ability to deploy it in an air-gapped environment, is what makes the cybersecurity angle concrete rather than marketing language. The License: Not Apache 2.0 — Read It First Mistral Large 3, released in December 2025, shipped under Apache 2.0 with no commercial restrictions. Large 4 ships under the Mistral Ecosystem Open-Weight License, a modified MIT license with one specific carve-out: operating Large 4 as a service that third parties can steer requires a separate commercial agreement with Mistral. Building a product where your customers interact with or configure the model directly falls into that category. Internal deployments — a bank running it for internal agents, a hospital in a sovereign datacenter, a team embedding it in an on-premise product — are fully permitted without a commercial license. Mistral’s release post https://mistral.ai/news/mistral-large-4-oss/ frames this as enabling sovereign AI: full control over your data and deployment infrastructure, within the permitted use cases. If you are building a developer-facing product where users invoke the model, read the license on Hugging Face https://huggingface.co/mistralai/Mistral-Large-4-1T-A52B/blob/main/LICENSE before you ship. The distinction between “I run it internally” and “my customers run it through me” is the line that matters. The Verdict Mistral Large 4 is the largest open-weight frontier model available, and it is real: weights are public, hardware requirements are documented, vLLM support is live, and the cybersecurity benchmarks are competitive with closed models that do not allow access to their weights at any price. For teams with the infrastructure to run it — enterprises with on-prem GPU clusters, security teams needing air-gapped deployment, organizations in regulated industries — this is a meaningful development. For individual developers, it is a landmark to watch rather than deploy today. The gap between open-weight and proprietary frontier AI narrowed on October 29th. It did not close entirely, but it narrowed in ways that matter for anyone building systems where data sovereignty is not optional.