Mistral发布1万亿参数大模型Mistral Large 4

Mistral AI公司发布了新一代大模型Mistral Large 4（代号"Le Chonk"），这是一个拥有1万亿参数、49亿活跃参数的原生多模态模型13。该模型采用混合专家架构2，配备1.6亿视觉编码器2，支持100万token上下文窗口2。根据Mistral，这一模型在欧洲和美国的聚合基准中表现为最佳的开源模型1。

在关键领域的性能方面，Mistral Large 4在网络防御、制造业和金融等工作负载上实现了开源模型中的最先进水平13。在网络安全基准测试中，该模型在Artificial Analysis Cyber Index上排名全球前五，在开源模型中领先3，在Cybench中解决了93%的挑战3，在漏洞复现和修补测试中得分82%3。此外，该模型在视觉定位上超越了闭源前沿模型1，支持160多种语言，包括欧盟所有官方语言3。

从技术效率看，Mistral Large 4仅使用了4000块NVIDIA GPU进行训练，相比中国竞争对手少2-3倍，比闭源竞争对手明显更少4。模型在Mistral欧洲数据中心使用NVIDIA GPU从头训练3，可通过Mistral Cloud基础设施从欧洲部署1。

该模型当前已通过API接口提供，定价为每百万输入token 0.68美元（标准版）或0.07美元（缓存输入），每百万输出token 2.09美元2。Mistral计划在10月底前发布开源权重版本13。根据Mistral副总裁Pierre Stock的说法，公司将与可信的合作伙伴和政府合作，"确保开源权重可用于防御，但不能进行恶意攻击"4。

Mistral AI has released Mistral Large 4 (also known as "Le Chonk"), a multimodal model featuring one trillion parameters with 49 billion active parameters, now available via API 123. The model incorporates a 1.6 billion-parameter vision encoder and supports a context window of one million tokens 2. Trained using 3,800 NVIDIA Grace Blackwell GPUs at Mistral's European data center 3, the model demonstrates advanced capabilities across critical enterprise workloads.

On the Artificial Analysis Cyber Index, Mistral Large 4 ranks in the global top five and leads among open-source models, achieving an 82% score on vulnerability reproduction and patching tests 3. The model solves 93% of challenges in Cybench and scores 61.7% on DeepSWE v1.1 3. It achieves 59.9% on AutomationBench, which evaluates 657 commercial workflows 3, and resists 93.3% of attacks on the Lakera B3 AI Security Benchmark 3. The model supports over 160 languages, including all official European Union languages 3, and excels in coding, agentic workflows, and multimodal understanding 3.

Pricing for Mistral Large 4 is set at $0.68 per million input tokens for standard usage and $0.07 per million tokens for cached input, with output priced at $2.09 per million tokens 2. The model is currently accessible through Mistral Studio's preview API 3. Mistral plans to release open-source model weights by the end of October 1, following completion of security testing 4. According to Vice President Pierre Stock, "We will work with trusted partners and governments to ensure open-source weights are available for defense purposes, but not for malicious attacks" 4. The company positions the model as a third path distinct from American closed-source and Chinese open-source alternatives 4.