Mistral Large 4
A flagship open-source large language model released by Mistral, built on a Mixture-of-Experts (MoE) architecture. It has approximately one trillion total parameters, with roughly 50 billion parameters activated per token. The model supports multimodal input and an extended context window of up to one million tokens.
Core Facts
Mistral Large 4 is priced at approximately $0.68 per million input tokens and $2.09 per million output tokens, with a 50% discount for the first two weeks.
Mistral Large 4 is the largest open-source model Mistral has released to date
Mistral Large 4 scored only 38 points on the Artificial Analysis composite intelligence index, ranking sixth among open-source models after excluding closed-source models.
Timeline (last 90 days)
Per-token price does not equal per-answer price: Gemini and Mistral have cheap per-token prices, but with an average of as many as 38 steps per question, their per-answer cost is comparable to or even higher than the currently used model.
In the speed test, Claude Sonnet and Haiku were the fastest and most stable, never exceeding 3 minutes; with Mistral and Gemini, about one in every four questions took over 10 minutes, and some approached 25 minutes.
Mistral Large 4 was excluded from the official ranking because it did not offer zero data retention (ZDR) at the time, and was tested only for reference.
Even if the model weights are open, self-hosting still requires a considerable scale of hardware investment.
ML4's training was completed in Europe, using about 3,800 NVIDIA Grace Blackwell GPUs.
The ML4 model weights are planned to be released in phases by the end of October 2026.
Mistral has offered a public preview of ML4 through Mistral Studio.
ML4's final license terms, deployment costs, and specific operational thresholds are currently not yet clear, and these details will only surface after the weights are actually released.
ML4 is a model with 1 trillion total parameters and 52 billion active parameters, continuing Mistral's路线 on large-scale sparse architectures.
The company positions it as a frontier open-weights model for text and multimodal work, with application scenarios covering fields such as cybersecurity, finance, and law.
40 more timeline events
All Facts (20)
Mistral Large 4 is priced at approximately $0.68 per million input tokens and $2.09 per million output tokens, with a 50% discount for the first two weeks.
80%VerifiedMistral Large 4 is the largest open-source model Mistral has released to date
75%VerifiedMistral Large 4 scored only 38 points on the Artificial Analysis composite intelligence index, ranking sixth among open-source models after excluding closed-source models.
70%VerifiedMistral Large 4 was trained using 3,800 NVIDIA Grace Blackwell GPUs in its own data centers, and the public preview also runs on the same infrastructure.
70%VerifiedMistral Large 4 uses a MoE architecture with 1 trillion total parameters and 49 billion active parameters, is natively multimodal, and has a context window of 512,000 tokens.
65%VerifiedMistral Large 4 narrowly outperformed GPT-6 Astra on visual grounding tasks with a score of 42% versus 41%.
65%VerifiedMistral Large 4 is a hybrid instruction-and-reasoning MoE model with a total of 1 trillion parameters, activating only approximately 49 billion parameters during inference.
65%VerifiedOn the Artificial Analysis Cyber Index, Mistral Large 4 scored 50, tying for first place with GLM 5.3 Flash.
65%VerifiedOn the DeepSWE 1.1 coding benchmark, Mistral Large 4 Preview scored 62, ranking second, behind only Kimi K3 paired with Kimi Code CLI at 68.
65%VerifiedMistral released Mistral Large 4, codenamed LeChunk, a trillion-parameter, natively multimodal MoE model with approximately 49 billion active parameters.
85%VerifiedMistral has launched its new flagship model Mistral Large 4, codenamed 'Le Chonk'.
80%VerifiedMistral has released Mistral Large 4, a one-trillion-parameter, natively multimodal MoE model with approximately 49 billion active parameters, with open-source weights to be released at the end of the month.
75%UnverifiedMistral Medium 3.5是Mistral于2025年4月发布的旗舰模型，专为智能体和编码工作场景优化
90%Partially VerifiedMistral Large 4 adopts a sparse Mixture of Experts (MoE) architecture, with 49 billion (49B) parameters actually activated during each inference.
75%Partially VerifiedMistral Large 4 is currently still in preview, and the model weights will be officially released later this month
65%UnverifiedMistral Large 4 was trained entirely from scratch in Mistral's own European data centers using 3,800 Grace Blackwell GPUs, with no distillation performed.
60%UnverifiedFrance's Mistral has released Mistral Large 4, scoring 38 points in a third-party intelligence index evaluation.
60%UnverifiedIn the speed test, Claude Sonnet and Haiku were the fastest and most stable, never exceeding 3 minutes; with Mistral and Gemini, about one in every four questions took over 10 minutes, and some approached 25 minutes.
50%UnverifiedPer-token price does not equal per-answer price: Gemini and Mistral have cheap per-token prices, but with an average of as many as 38 steps per question, their per-answer cost is comparable to or even higher than the currently used model.
50%UnverifiedMistral Large 4 was excluded from the official ranking because it did not offer zero data retention (ZDR) at the time, and was tested only for reference.
50%