Mistral AI Releases 1-Trillion-Parameter Large 4 Model with Open Weights

Paris-based Mistral AI officially released the full open weights for its latest flagship model, Mistral Large 4, on October 29. Known internally by the moniker Le Chonk, the architecture represents a significant expansion in scale, featuring approximately 1 trillion total parameters and a mixture-of-experts design that utilizes roughly 52 billion active parameters per token.

The release follows a three-week post-preview period during which the development team continued reinforcement learning cycles. This extended training phase resulted in a notable performance increase, with the model’s Terminal-Bench 4.0 score rising from 28.3 percent at the time of the October 6 preview to 41 percent upon the public release. The architecture incorporates a 1-million-token context window and a 1.6-billion-parameter Pixtral vision encoder to support native multimodal capabilities.

Technical specifications for the model include 128 routed experts, with four active experts per token alongside one shared expert. The design utilizes MLA attention mechanisms paired with a sparse-attention indexer modeled after DeepSeek architectures. Reasoning effort is exposed through three distinct modes—low, medium, and high—with the high-effort setting serving as the default configuration for complex tasks.

Deployment flexibility is supported through multiple checkpoint formats, including FP8 and NVFP4, with community-driven GGUF quantizations available via llama.cpp and Unsloth. Initial testing indicates that 2-bit quantizations can be hosted on a single Apple Silicon Ultra machine equipped with 512 gigabytes of unified memory. Conversely, production environments utilizing vLLM with the NVFP4 checkpoint require the high-throughput capabilities of Blackwell-generation GPUs.

The model demonstrates competitive performance across several specialized benchmarks, particularly in coding and security domains. On the FrontierCode 2.0 benchmark, the model achieved a score of 19.7 percent, while it reported 97.1 percent attack resistance on the B3 agent-security benchmark. These metrics have attracted early endorsements from organizations including Booz Allen, Cloudflare, and Orange Cyberdefense.

Efficiency gains remain a core component of the release, as the model demonstrates a 30 percent reduction in output tokens compared to the preview version. On the Artificial Analysis Coding Agent Index, the model scores 48 percent, effectively tying with GLM-5.3 while utilizing 47 percent fewer tokens per task. Such improvements suggest that the continued reinforcement learning cycles prioritized inference optimization alongside raw reasoning capabilities.

Despite these gains, independent assessments provide a more nuanced view of the model’s standing within the global landscape. Data from the Artificial Analysis Intelligence Index v5 places the model at a score of 26, marking it as the highest-performing Western open-weight model while trailing several Chinese-developed alternatives in aggregate rankings. The model’s Cyber Index score of 51 percent similarly places it behind DeepSeek and MiMo in current open-weight comparisons.

The release also clarifies the commercial framework surrounding Mistral’s ecosystem. The model is distributed under a modified MIT license that permits free usage, modification, and fine-tuning for most applications. However, organizations intending to offer the model as a service are required to negotiate a separate commercial license, a distinction that continues to fuel industry debates regarding the precise definition of open-source software versus open-weight distribution.

Pricing adjustments accompany the release, with API costs now set at 0.70 dollars per million input tokens and 2.10 dollars per million output tokens. This represents a roughly 50 percent reduction from the pricing structure established during the preview phase. The shift in pricing and the continued improvement of the model post-announcement signal a strategic effort to capture market share among enterprise developers.

European sovereignty remains a central theme of the release, as the availability of domestically trained high-parameter models provides a strategic alternative for regional governments and enterprises. As procurement policies increasingly prioritize jurisdiction-bound suppliers, the ability to deploy a 1-trillion-parameter model locally offers a distinct advantage for sensitive data processing. Future adoption will likely depend on how effectively the model maintains its efficiency gains while scaling across diverse production environments.