Mistral AI has officially released Mistral Large 4 (ML4), internally nicknamed "Le Chonk," to the public in a preview format. This new flagship model is a sophisticated multimodal Mixture of Experts (MoE) architecture. Its architecture encompasses an impressive 1.05 trillion parameters in total, yet only 49 billion are active during each token's processing. It also integrates a vision encoder comprising 1.6 billion parameters and provides an extensive context window capable of handling 1 million tokens, according to its official documentation.
The development of ML4 was an intensive process, involving training from scratch on 3,800 NVIDIA Grace Blackwell graphics processing units within Mistral's own European data centers. The API for Mistral Large 4 is now accessible, with pricing set at $1.36 per 1 million input tokens and $4.18 for every 1 million output tokens. While the API is live, the model's weights are not expected to be released until late October, meaning self-hosting capabilities will be available at a later date.
Advanced Architecture and Efficiency
ML4 distinguishes itself with a hybrid instruct-and-reasoning MoE design that natively accepts image input. A key architectural feature is its efficiency; approximately 4.7% of the total weights are activated per token, allowing a model of this scale to be served at competitive pricing. Although the entire 1.05 trillion parameters must reside in memory, the active parameter count directly influences compute costs rather than hardware expenditure. Further details regarding the expert count, top-k routing mechanisms, and layer configurations are anticipated with the upcoming release of the model weights. The training dataset for ML4 encompassed more than 160 languages, including every official language used within the European Union.
Performance Across Key Domains
Mistral AI highlights ML4's strong performance, particularly in cybersecurity. The model achieved a 93% score on Cybench and an 82% on CyberGym-E2E, positioning it among the global top five on the Artificial Analysis Cyber Index. A notable claim by Mistral is that several competing closed frontier models reportedly score near zero on CyberGym-E2E due to refusing the task outright. This highlights ML4's capability in tasks considered standard defensive work, such as reproducing vulnerabilities.
In the domain of agentic coding, ML4 recorded 61.7% on DeepSWE v1.1, 59.4% on SWE-Atlas-QnA, and 28.3% on Terminal-Bench 4.0, resulting in a combined Artificial Analysis Coding Agent Index of 49.8%. Mistral clarified that these evaluations were conducted privately before the harness became public, awaiting independent reproducibility. A blind human evaluation carried out with Surge AI also placed ML4 Preview at 3.74 out of 5, ranking it second among five evaluated models. It outperformed GLM-5.3 (3.60) and Kimi K3 (3.59) but trailed Claude Opus 5 (4.22).
Regarding safety, ML4 demonstrated robust resistance, fending off 93.3% of attacks on Lakera's B3 benchmark and securing a score of 1.691 out of a maximum 2.0 on KORABench.
Competitive Positioning and Capabilities
When compared to other open-weight models, Mistral Large 4 stands out with its 1 million token context window and native image input capabilities, features it shares with some rivals like Kimi K3, but not others such as DeepSeek V4 Pro or GLM-5.3. While its total parameter count of 1.05T is substantial, some models like Kimi K3 (2.8T) and DeepSeek V4 Pro (1.6T) report higher aggregates, though active parameters often vary. ML4's API pricing for input and output tokens is competitive within the landscape of similar advanced models.
The preview API currently supports a range of functionalities, including advanced function calling, structured output generation, document question-answering, and batch processing. It also provides dedicated endpoints for building agents and managing conversations. Economically, cached input is priced at $0.14 per 1 million tokens, which significantly alters the cost dynamics for long-context agent loops.
This article is a rewritten summary based on publicly available reporting. For the original story, visit the source.
Source: MarkTechPost
