THE SHORT VERSION
- Mistral published the open weights of Mistral Large 4 on October 29, a 1-trillion-parameter mixture-of-experts model with about 52 billion active parameters and a 1M-token context window.Mistral 于 10 月 29 日发布 Mistral Large 4 的开放权重：万亿参数的混合专家模型，约 520 亿激活参数，支持 100 万 token 上下文。它的视觉编码器采用 Pixtral 风格，可处理不同尺寸与长宽比的图像，模型也支持原生函数调用与严格 JSON schema。
- The release license is a modified MIT license: free to use, modify, fine-tune and commercialize, except that businesses running a model-as-a-service offering need a separate commercial license.许可证是修改版 MIT：可自由使用、修改、微调和商业化，但经营模型即服务业务的企业需另行取得 Mistral 的商业许可。
- Mistral reports leading open-weight results in agentic coding and defensive cybersecurity, while noting the model was trained on 3,800 Nvidia Grace Blackwell GPUs in its own European data centers.Mistral 称该模型在智能体编程与防御性网络安全上领先开放权重模型，并说明它是在欧洲自有数据中心、用 3800 块英伟达 Grace Blackwell GPU 从零训练的。
The weights are out, after the preview预览之后，权重正式开放
Mistral published the weights of Mistral Large 4 after releasing the model in public preview earlier this month. The company's announcement says the release follows additional reinforcement learning aimed at coding and defensive cybersecurity, and that the gains had not yet plateaued. The model is available through Mistral's API and Studio, on OpenRouter, and as a download from Hugging Face, where the model card lists deployment recipes for vLLM and SGLang. A quantized NVFP4 checkpoint for Blackwell GPUs is also published, and the card tags the release as image-text-to-text.Mistral 在本月早些时候以公开预览形式发布后，又发布了 Mistral Large 4 的权重。公司的发布说明称，此次发布前又进行了针对编程与防御性网络安全的强化学习，且提升尚未见顶。模型可通过 Mistral 的 API 与 Studio、OpenRouter 使用，也可从 Hugging Face 下载；模型卡给出了 vLLM 与 SGLang 的部署方案，同时还发布了面向 Blackwell GPU 的 NVFP4 量化版本，并把该版本标注为图文输入到文本输出。
The model card describes a 1-trillion-parameter mixture-of-experts model with 49 billion active parameters per token, or 52 billion including embeddings and output layers. It has 128 routed experts with four active per token and one shared expert, accepts text and image input, produces text, and supports a 1-million-token context window. Its vision encoder is Pixtral-style and handles variable image sizes and aspect ratios, and the model supports native function calling and strict JSON schemas. Mistral says it trained the model from scratch on 3,800 Nvidia Grace Blackwell GPUs in its own data centers in Europe.模型卡描述这是一个万亿参数的混合专家模型，每个 token 约 490 亿激活参数，计入嵌入与输出层后约 520 亿。它包含 128 个路由专家，每个 token 激活 4 个，另有 1 个共享专家；可接受文本与图像输入、输出文本，支持 100 万 token 上下文。Mistral 称该模型是在欧洲自有数据中心、用 3800 块英伟达 Grace Blackwell GPU 从零训练的。
The license is almost MIT, with one condition许可证接近 MIT，但保留一个条件
The release is not unrestricted. Mistral published the model under what it calls the Mistral Ecosystem Open-Weight License, a modified MIT license. The company says the model is free to use, modify, fine-tune and commercialize, with one condition: a business that operates a model-as-a-service offering must obtain a separate commercial license from Mistral before using the model commercially.这次发布并非毫无限制。Mistral 以“Mistral 生态系统开放权重许可证”发布模型，这是修改版的 MIT 许可证。公司称模型可自由使用、修改、微调和商业化，但有一个条件：经营模型即服务业务的企业，在商业使用前必须另行取得 Mistral 的商业许可。
That condition separates two kinds of user. A company that downloads the weights to run them internally, or to build a product on top, falls under the open license. A company that resells access to the model as a service needs a deal with Mistral. The release does not say how Mistral would identify such a use, so enforcement rests on the licence terms rather than on a technical restriction. The distinction matters for cloud providers and inference startups more than for enterprises running the model in their own data centers.这个条件把用户分成两类。把权重下载到内部运行、或在其上构建产品的公司，适用开放许可证；把模型访问权作为服务转售的公司，则需要与 Mistral 单独签约。发布稿没有说明 Mistral 将如何识别这类使用，执行只能依靠许可证条款，而非技术限制。对云服务商和推理初创公司来说，这一区别比对自己在数据中心运行模型的企业更重要。
Mistral's numbers put it at the open-weight top for coding按 Mistral 的数据，它处在开放权重编程模型顶端
Mistral reports several benchmark results, all of which are the company's own summaries of third-party evaluations. On Cognition's private FrontierCode 2.0 benchmark, which tests whether a repository maintainer would merge an agent's change, Mistral Large 4 scores 19.7 percent, the company reports. On the Artificial Analysis Coding Agent Index, it scores 48 percent, effectively tied with GLM-5.3 while using 47 percent fewer tokens per task on average. On the Artificial Analysis Intelligence Index v5 the model scores 26, which Mistral describes as the highest among western open-weight models; Chinese open models score higher on that aggregate measure.Mistral 报告了多项基准结果，这些结果都是公司对第三方评测的自行汇总。在 Cognition 的私有基准 FrontierCode 2.0 上——该基准考察仓库维护者是否愿意合并智能体提交的改动——Mistral 报告 Mistral Large 4 得 19.7%。在 Artificial Analysis 的编程智能体指数上，它得 48%，与 GLM-5.3 基本持平，但每个任务平均少用 47% 的 token。在 Artificial Analysis 智能指数 v5 上，该模型得 26 分，Mistral 称这是西方开放权重模型中的最高分；中国开放模型在该综合指标上更高。
On Terminal-Bench 4.0, the company says the model improved by 13 percentage points over its preview to 41 percent, leading open-weight models while cutting output tokens by 30 percent. In Surge Code's blind pairwise evaluations, graders preferred Mistral Large 4 over GLM-5.3, Kimi K3 and DeepSeek V4.1 Flash, and preferred Claude Opus 5.5 over it. The announcement states that ranking itself, including the model it loses to. These are vendor-reported figures; the benchmarks are third-party, but the selection and presentation come from Mistral.在 Terminal-Bench 4.0 上，公司称该模型比预览版提升 13 个百分点至 41%，领先开放权重模型，同时减少 30% 的输出 token。在 Surge Code 的盲评两两对比中，评审更偏好 Mistral Large 4 而非 GLM-5.3、Kimi K3 与 DeepSeek V4.1 Flash，但更偏好 Claude Opus 5.5。发布稿把这一排名本身也写了出来，包括它输给的那款模型。这些都是厂商报告的数字；基准本身来自第三方，但选择与呈现方式来自 Mistral。
Defensive security is the other pitch防御性安全是另一半卖点
Mistral is marketing the release to security teams as much as to developers. It reports a score of 51 percent on the Artificial Analysis Cyber Index, which measures finding, reproducing, validating and patching vulnerabilities in a codebase, and says the model handles those defensive tasks without unnecessary refusals. It also reports an average refusal rate of 96.6 percent on malicious cyber requests across three jailbreak and harm benchmarks.Mistral 把这版模型同时推销给安全团队。它报告在 Artificial Analysis 网络安全指数上得 51%，该指数衡量在代码库中发现、复现、验证并修补漏洞的能力；公司称模型处理这类防御任务时不会出现不必要的拒答。它还报告，在三项越狱与危害基准的恶意网络安全请求上，平均拒答率为 96.6%。
The company says it monitored live traffic after the preview and observed the model refusing black-box attacks against live targets while seeing potential for uplift in real attacks, and that it is adjusting guardrails. ENISA, the EU cybersecurity agency, is testing the model to identify vulnerabilities, audit code and analyse defensive logs, and Booz Allen, Cisco and Cloudflare also provided statements in the announcement. Enterprise design partners, including Orange Cyberdefense, are getting preview access to a specialized variant for penetration testing, vulnerability assessment and red-teaming, subject to vetting and contracts that restrict testing to owned or authorized systems. Mistral also says customers can tailor the model with proprietary security data through Mistral Forge.公司称在预览后监测了真实流量，观察到模型会拒绝对真实目标发起黑盒攻击，但也看到被用于真实攻击的潜在提升，并正在相应调整防护栏。欧盟网络安全局 ENISA 正在测试该模型，用于发现漏洞、审计代码和分析防御日志；Booz Allen、Cisco 与 Cloudflare 也在发布稿中提供了声明。包括 Orange Cyberdefense 在内的企业设计伙伴，将获得针对渗透测试、漏洞评估和红队工作的专用变体预览权限，但须经过审查并签署合同，测试仅限自有或获授权的系统。Mistral 还表示，客户可通过 Mistral Forge 用自有安全数据定制模型。
Finance, browsing and agent security金融、浏览与智能体安全
Beyond coding, the company reports a Finch score of 73.3 percent for finance workflows, second among open models to Kimi K3, and 84.5 percent on BrowseComp for web browsing. On the B3 Agent Security Benchmark, it says the model resisted 97.1 percent of attacks, up from 93.3 percent in the preview, which the company presents as evidence that organizations can host it without additional tuning. The demos are published in a gallery that runs on the Mistral API with a small shared free budget for the first week, and each one is a company-selected task rather than a measured failure rate.在编程之外，公司报告金融工作流的 Finch 得分 73.3%，在开放模型中仅次于 Kimi K3；网页浏览的 BrowseComp 得 84.5%。在 B3 智能体安全基准上，它称模型抵御了 97.1% 的攻击，高于预览版的 93.3%，公司把这作为企业无需额外调优即可自托管的证据。演示被收录进一个画廊，第一周通过 Mistral API 提供少量共享免费额度；每个演示都是公司挑选的任务，而不是测得的失败率。
The announcement also lists demonstrations rather than benchmarks: an agent that triages vulnerability reports and iterates on a patch until a canary exploit fails, a generated 3D map of about 19,000 satellites and debris pieces, and a quantum chemistry solver checked against a standard tool. These are company-selected examples, and they show what the model can be asked to do rather than how reliably it does it.发布说明还列出了一些演示而非基准：一个智能体阅读漏洞报告、修补代码并迭代到金丝雀漏洞利用失效；生成一张约 1.9 万颗卫星与碎片的地球轨道 3D 地图；以及一个量子化学求解器，结果与标准化学工具对照。这些是公司挑选的例子，说明模型可以被要求做什么，而不是它有多可靠。
What it costs to use, and what running it takes使用成本，以及运行它需要什么
On Mistral's API, the model is priced at $0.70 per million input tokens, $0.07 per million cached input tokens, and $2.10 per million output tokens. Microsoft is the first cloud provider to offer it, through Microsoft Foundry, and Nvidia hosts endpoints and publishes an optimized NIM container. The model card also notes a quantized NVFP4 checkpoint for Blackwell GPUs and an embedded speculative decoding head that Mistral says can raise throughput up to 1.9 times.在 Mistral 的 API 上，该模型定价为每百万输入 token 0.70 美元、每百万缓存输入 token 0.07 美元、每百万输出 token 2.10 美元。微软是首家提供该模型的云厂商，通过 Microsoft Foundry 上线；英伟达提供托管端点并发布优化过的 NIM 容器。模型卡还提到面向 Blackwell GPU 的 NVFP4 量化版本，以及内嵌的推测解码头，Mistral 称可把吞吐量提高至多 1.9 倍。
Self-hosting is the harder part. The model card's vLLM recipe uses tensor parallelism across eight GPUs, and a full-precision trillion-parameter model needs substantial memory. The license makes the weights available, and the hardware still decides who can serve them at scale. The published vLLM recipe assumes an eight-GPU node, which is the practical requirement for serving the checkpoint. That gap between an open license and a practical deployment is the part a reader should keep in view when the release is described as open. Related coverage includes the Mistral Responses API preview, Meta's open-weight Muse Spark, and Models. Another open-weight release, Reflection's Beam.自托管是更困难的部分。模型卡的 vLLM 方案使用跨 8 块 GPU 的张量并行，而全精度的万亿参数模型需要大量显存。许可证让权重可得，硬件仍决定谁能大规模提供服务。模型卡给出的 vLLM 方案假设使用 8 块 GPU 的节点，这是服务该检查点的实际要求。当这次发布被描述为“开放”时，开放许可与实际部署之间的这道差距，是读者应当留意的部分。 相关报道包括 Mistral Responses API 预览、Meta 的开放权重 Muse Spark，以及模型动态。 另一次开放权重发布——Reflection 的 Beam。
Sources & reporting limits来源与核查范围
Follow Models to find related stories on your next visit.关注模型动态，下次访问时阅读相关文章。
