Coding performance: Reported Terminal-Bench 4.0 performance rose from 28% to 41%, a 13-percentage-point improvement following additional reinforcement learning.1011 Its reported Artificial Analysis Coding Agent Index score is 48%, tied with GLM-5.3 while using 47% fewer tokens per task.11 These are reported benchmark results—not evidence of universal coding superiority; the available excerpts do not verify the exact three-week training duration, the 30% Terminal-Bench token reduction, or a numerical comparison with Anthropic’s Mythos 5.1.

Architecture and compute: Official documentation describes a multimodal mixture-of-experts model with 1.05 trillion total parameters, 52 billion active parameters, and a 1.6-billion-parameter vision encoder.1 The Hugging Face description distinguishes 49 billion active parameters per token from 52 billion when embeddings and the output layer are included.[12 from previous search] Reasoning settings are reported as low, medium, and high, with high the default.10 Mistral says training used 3,800 Nvidia Grace Blackwell GPUs in its European data centers.15

Availability and licensing: A Hugging Face release commit contains the “Mistral Ecosystem Open-Weight License,” explicitly described as a modified MIT license.7 Secondary reporting says its additional commercial restrictions target model-as-a-service resellers, rather than self-hosters, fine-tuners, or end-product builders.12 Insufficient evidence in the available excerpts to specify the exact obligations imposed on hosted providers or verify Microsoft Foundry availability.

Cybersecurity and safety: Secondary reporting gives a 51% Cyber Index score and a claimed 96.6% refusal rate for malicious cyber requests.8 Those figures should be treated as benchmark and evaluation claims, not a guarantee against misuse. The available excerpts do not establish the test methodology, Cisco/CrowdStrike/Cloudflare support, or precisely how guardrails changed after evidence of real-world attack assistance.

European competition: The release gives users an additional European open-weight option that Mistral positions against Chinese rivals and closed frontier models.914 Analytically, the important distinction is control over a downloadable checkpoint versus dependence on continued access to a provider’s API—not nationality alone.

Efficiency: Matching a competitor’s reported coding score with fewer tokens suggests a potentially useful efficiency advantage.11 It does not, by itself, establish lower total deployment costs; that requires hardware, latency, pricing, and workload measurements.

Dual-use risk: Defensive cybersecurity capability and malicious-request refusal are separate dimensions. A strong defense benchmark should not be read as proof that releasing the weights is risk-free.

Claims not verified: Insufficient evidence for Pierre Stock’s specific comments about irrevocable access and trust in Chinese models; Anthropic’s alleged nearly three-week suspension following a June export-control directive; Mistral’s approximately $6 billion funding total; a Samsung-led €3 billion Series D above a €21 billion valuation; or Rafe Pilling’s specific remarks about dual-use risks and nationalistic framing. These details should not be presented as established facts from the excerpts supplied.

Coding performance: Reported Terminal-Bench 4.0 performance rose from 28% to 41%, a 13-percentage-point improvement following additional reinforcement learning.1011 Its reported Artificial Analysis Coding Agent Index score is 48%, tied with GLM-5.3 while using 47% fewer tokens per task.11 These are reported benchmark results—not evidence of universal coding superiority; the available excerpts do not verify the exact three-week training duration, the 30% Terminal-Bench token reduction, or a numerical comparison with Anthropic’s Mythos 5.1.

Architecture and compute: Official documentation describes a multimodal mixture-of-experts model with 1.05 trillion total parameters, 52 billion active parameters, and a 1.6-billion-parameter vision encoder.1 The Hugging Face description distinguishes 49 billion active parameters per token from 52 billion when embeddings and the output layer are included.[12 from previous search] Reasoning settings are reported as low, medium, and high, with high the default.10 Mistral says training used 3,800 Nvidia Grace Blackwell GPUs in its European data centers.15

Availability and licensing: A Hugging Face release commit contains the “Mistral Ecosystem Open-Weight License,” explicitly described as a modified MIT license.7 Secondary reporting says its additional commercial restrictions target model-as-a-service resellers, rather than self-hosters, fine-tuners, or end-product builders.12 Insufficient evidence in the available excerpts to specify the exact obligations imposed on hosted providers or verify Microsoft Foundry availability.

Cybersecurity and safety: Secondary reporting gives a 51% Cyber Index score and a claimed 96.6% refusal rate for malicious cyber requests.8 Those figures should be treated as benchmark and evaluation claims, not a guarantee against misuse. The available excerpts do not establish the test methodology, Cisco/CrowdStrike/Cloudflare support, or precisely how guardrails changed after evidence of real-world attack assistance.

European competition: The release gives users an additional European open-weight option that Mistral positions against Chinese rivals and closed frontier models.914 Analytically, the important distinction is control over a downloadable checkpoint versus dependence on continued access to a provider’s API—not nationality alone.

Efficiency: Matching a competitor’s reported coding score with fewer tokens suggests a potentially useful efficiency advantage.11 It does not, by itself, establish lower total deployment costs; that requires hardware, latency, pricing, and workload measurements.

Dual-use risk: Defensive cybersecurity capability and malicious-request refusal are separate dimensions. A strong defense benchmark should not be read as proof that releasing the weights is risk-free.

Claims not verified: Insufficient evidence for Pierre Stock’s specific comments about irrevocable access and trust in Chinese models; Anthropic’s alleged nearly three-week suspension following a June export-control directive; Mistral’s approximately $6 billion funding total; a Samsung-led €3 billion Series D above a €21 billion valuation; or Rafe Pilling’s specific remarks about dual-use risks and nationalistic framing. These details should not be presented as established facts from the excerpts supplied.