Mistral Large 4 weights released with NVFP4 support for Blackwell GPUs

The official weight release

Mistral AI has officially released the weights for Mistral Large 4, correcting several early rumors regarding its availability. While some reports suggested the license was missing or the checkpoints were unavailable, the files are now live in the repository. The release includes a primary FP8 checkpoint at 1,060.6 GB and a smaller NVFP4 version at 610.1 GB.

The model is a Mixture-of-Experts (MoE) architecture featuring 1.05 trillion parameters. In this setup, only 49 billion to 52 billion parameters are active per token, depending on whether you count the embeddings and output layers. It also includes a 1.6 billion parameter vision encoder capable of processing images up to 1,540 pixels.

As OrcaRouter reported, the release arrived slightly later than some press outlets predicted. This delay meant that much of the early coverage was based on projections rather than the actual files. The current repository now provides the definitive specifications for the model's architecture and size.

Hardware and memory optimization

The 610.1 GB NVFP4 checkpoint is specifically designed to solve a memory bottleneck for large-scale inference. A model with 1.05 trillion parameters is too large to fit on standard hardware without significant compression. On an 8-GPU Blackwell node, the FP8 version leaves almost no room for the KV-cache, which is necessary for handling long context windows.

By using the NVFP4 quantization for the Mixture-of-Experts weights while keeping the linear layers in FP8, Mistral allows the model to fit on a single node with enough room for the 1 million token context. This is a mixed-precision serving stack rather than a simple downsampling of the entire model. It is specifically optimized for NVIDIA Blackwell components like the B200 and GB200.

The memory efficiency is also achieved through aggressive attention sharing. The configuration uses a Sparse-Indexer with a Top-k of 2,048 over 32 heads and KV-sharing across most of the 72 layers. These techniques ensure that the 1 million token context is reachable without requiring a terabyte of memory for the cache.

Serving recipes and DSpark

Unlike many open-weight releases where the community must figure out the best way to run the model, Mistral provided specific serving instructions. They released distinct paths for both vLLM and SGLang frameworks. These instructions include specific flags for tensor parallelism, memory utilization, and specific parsers for tool calling and reasoning.

One notable feature is the inclusion of a DSpark draft head within the 610 GB checkpoint. Usually, speculative decoding requires running a second, smaller model alongside the large one, which costs extra memory. Because this draft head is already embedded in the weights, users can enable speculative decoding via a simple configuration string rather than a second deployment.

The model also uses a YaRN block to expand the native RoPE window. This allows the model to reach the advertised 1,048,576 token context length through extrapolation. Knowing this is an extrapolation helps developers understand that the model may not have seen every possible 900,000-token document during its initial training phase.

Pricing and performance data

The release of the weights has also shifted the pricing landscape for the model's API. The initial announcement listed prices of $1.36 for input and $4.18 for output per million tokens. However, the current documentation shows a significant drop to $0.70 for input and $2.10 for output, bringing the price closer to the original preview rates.

Performance benchmarks provided by Mistral suggest the model is highly capable in coding and reasoning tasks. It reportedly achieves a 41% score on Terminal-Bench 4.0 and a 48% score on the AA Coding Agent Index. These numbers are significantly higher than the preview versions, though they are reported by Mistral itself rather than independent third parties.

Independent analysis from Artificial Analysis shows a different picture for the hosted API. They measured a Terminal-Bench 4.0 score of 25%, which is 16 points lower than Mistral's reported figure. This discrepancy likely stems from different configurations or specific reasoning settings used during the evaluation. It highlights the importance of checking specific benchmarks before making production decisions.

Licensing and next steps

The model is released under the Mistral Ecosystem Open-Weight License, which Mistral describes as a modified MIT license. However, there is a critical distinction for businesses regarding