最新开放成果（#25）：DeepSeek V4.1 Flash 1017、美国本土模型与中美新监管政策

今年秋季，开放权重（open-weight）AI 的地缘政治博弈再度升温，正如我们所预期的那样，所有 AI 系统（无论是开源还是闭源）的潜在风险都在快速上升。过去一周涌现出诸多新闻，表明中美两国政府都在为针对开放模型施加某种监管进行布局。

1. 中国商务部（MOFCOM）发布了一份政策草案提案（目前尚非正式法律），要求前沿 AI 模型以及其他部分技术在进行国际共享前必须获得政府的前置审批。我们认为外界对此反应有些过度——因为该文件明确指出，开源技术将设立自动审批机制——但地缘政治圈内的许多人认为这是一件大事，因为它提出了一个未来可用于监管中国 AI 产业的监管框架。

在这一议题上，其他我们信任的消息人士表示，该提案可能更多是为了防止类似中国初创公司 Manus 出售给西方公司这类情况的再次发生，而非针对前沿 AI 模型本身。无论如何，围绕 AI 的监管齿轮已经开始转动，开放权重模型也绝无法独善其身。

2. 美国国土安全部（DHS）发布了一项针对 DeepSeek 的禁令，限制 DeepSeek 作为一家公司及其模型作为技术平台，参与联邦民用采购供应商名录。目前尚不清楚这一禁令将如何影响微调模型或基于 DeepSeek 输出训练的模型（例如，Nvidia 的 Nemotron 模型长期以来在其训练配方中严重依赖 DeepSeek 模型——它们是否也会被禁？），但这是一项意料之中的举措，并进一步强化了构建本土开放模型生态系统的必要性。

中国其他模型厂商（如 GLM 的开发者智谱 AI）早已被列入美国商务部/工业和安全局（BIS）的实体清单，从而限制了其获取美国技术的途径。

DeepSeek 紧接着上述新闻，公开发布了其内部推理技术栈 DeepInfer，对于其自身模型而言，该工具的运行效率很可能远超现有的 vLLM/SGLang 等开源工具。相关估算显示，DeepSeek 借助这套推理栈销售 Token 时，毛利率最高可达 90%。众所周知，顶级 AI 实验室内部都拥有推理效率更高的自研工具，而这是我们首次看到顶尖实验室将其对外公开。推理服务提供商将迅速跟进采用这些工具，这将推动整个生态系统中 DeepSeek 模型的调用价格进一步下调（或提升利润率）。尽管它不是开源模型或数据集，但我们认为该发布是过去几个月中最重要的开源 AI 成果之一。

截至撰写本文时，传闻 Kimi K3.1 即将发布，其相关元数据已开始在公共代码库中出现。令人振奋的时代！

您可以在我们的 Artifacts Hub 上查看本期邮件中提到的模型以及往期所有成果，其中包含深度点评和其他生态元数据。

浏览 Artifacts Hub

• MiMo-V2.6-Pro-RL 由 XiaomiMiMo 提供：MiMo 的更新版本，沿用了与 MiMo V2.5 相同的基座，但在强化学习（RL）训练上投入了更多资源。他们还发布了一份深度技术报告以及部分所用环境，这种开放程度在如此规模下十分罕见。基于该配方打造的最终模型 V2.6-Pro 目前在 Artificial Analysis 排行榜上高居开源模型榜首，其大规模 RL 训练还配备了一个有趣的实时训练仪表盘。数月来我们一直是小米 AI 的拥趸，如今他们的努力正在开花结果。

• DeepSeek-V4.1-Flash-1017 由 deepseek-ai 提供：自上一期 Artifacts 以来，这头巨鲸并未只推出一款模型，而是连续发布了两次采用新尺寸（552B-A16B）的模型迭代，并且支持图像输入。该模型凭借高性价比获得了极高且实至名归的关注度。它同样遵循 MIT 许可证，延续了 DeepSeek 自身确立的开源传统。在当前各类模型许可证日益趋向定制化和排他性的背景下，这一点尤为瞩目。下方是该模型在当前备受青睐的 Agent 软件工程基准 DeepSWE 上的性能与价格帕累托前沿对比，DeepSeek 在该榜单上占据主导地位。

• clef 由 Cloudflare 提供：自从初创公司 Typesafe 提出决策模型（Decision models）概念后，这一方向便风靡一时。简而言之，它们是能够输出少量 Token 及其概率的通用分类模型。其高度的实用性促成了迅速的普及应用，并催生了包括 Clef 在内的大量开源微调模型。

• NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-NVFP4 由 nvidia 提供：Nemotron 家族的中型模型终于迎来了期待已久的更新。相对于其体积而言，该模型性能强劲，且支持在本地运行（前提是配备高性能笔记本电脑或 RTX Spark）。按照惯例，NVIDIA 开源了绝大部分训练数据，并附带了一份深度技术报告。

• Beam-1.0-501B-A23B 由 reflection 提供：Reflection 团队万众瞩目的开源模型，成为目前采用宽松许可证（Apache 2.0）发布的最强美国开源模型。尽管 Beam 并非市面上最聪明的开源模型，但其 Token 效率令人印象极其深刻（仅次于下文的 GPT-6-Astra），这充分展现了其成熟完善的后训练（post-training）技术栈。

The geopolitics of open-weight AI are accelerating again this fall, as we would expect with the risks of all AI systems (open and closed) rising rapidly. We’ve seen many news items in the last week showing both the U.S. and Chinese governments positioning themselves for some controls on open models.

1. MOFCOM, the Ministry of Commerce of the People’s Republic of China, released a policy proposal (it’s not yet a law) that would require frontier AI models and some other technologies to get pre-approval from the government before they can be shared internationally. We found the reaction to this to be overblown – as the document states that open-source technologies will be set up for automatic approvals – but many in geopolitical circles say this is a big deal because it proposes a regulatory framework that could be used in the future to regulate China’s AI industry.

Others we trust on this issue say this proposal could be more about preventing issues like the sale of Chinese startup Manus to Western companies, rather than being about frontier AI models. Either way, the wheels of control are moving around AI, and open-weight models are not free from any influence.

2. The Department of Homeland Security (DHS) released an order on DeepSeek, which restricts DeepSeek itself, as a company, and DeepSeek’s models, as a technology platform, from being a vendor in federal civilian procurement. It is not clear how this will relate to finetuned models or models trained on the outputs of DeepSeek (e.g., Nvidia’s Nemotron models have long relied heavily on DeepSeek models in their training recipes – will they be banned?), but this is an expected action and reinforces the need for a domestic open model ecosystem.

Other model makers in China, such as Zhipu AI, makers of GLM, are already on the Entity List under Commerce/BIS, which restricts their access to American technology.

DeepSeek followed the above news up by releasing their internal inference stack publicly, DeepInfer, which is likely far more efficient than existing open-source tools like vLLM/SGLang for their models. Estimates point to DeepSeek making up to 90% gross margin when selling tokens on this inference stack. It is well-known that the top AI labs have much more inference-efficient tools internally, and this is the first time we’ve seen a leading lab share them. Inference companies will be quick to adopt these tools, which will drive prices for DeepSeek’s models down across the ecosystem (or margins up). While not an open model or dataset, we consider this release to be one of the bigger open-source AI artifacts in the last few months.

Share

As of writing this, Kimi K3.1 is rumored to be released very soon, with metadata associated with it starting to appear in public libraries. Exciting times!

You can view the models in this email, and all previous issues, on our Artifacts Hub. It includes commentary and other ecosystem metadata.

Browse the Artifacts Hub

• MiMo-V2.6-Pro-RL by XiaomiMiMo: an update for Mimo, which reuses the same bases as MiMo V2.5, but used a lot more resources on RL training. They also release an in-depth tech report, as well as some of the used environments, which is rare at this scale. The final model from this recipe, V2.6-Pro, is the top open model on Artificial Analysis’s ranking today, and the scaled RL runs came with a fun live dashboard of training. We’ve been fans of Xiaomi for months now, and their efforts are bearing fruit.

• DeepSeek-V4.1-Flash-1017 by deepseek-ai: Since the latest Artifacts, the whale has not released one, but two new model iterations with a new size (552B-A16B) which also support image inputs. The model has gained a lot of (well-deserved) attention for its performance at a small price. It also carries the MIT license, continuing the trend established by DeepSeek themselves. This is notable in times where licenses seem to become more bespoke and exclusionary. Below is the Pareto of performance vs price on DeepSWE, a favorite agentic software engineering benchmark today, which DeepSeek dominates.

• clef by Cloudflare: Decision models are all the rage after being introduced by the startup Typesafe. In short, they are general classification models that output few tokens with their probabilities. Its usefulness has led to quick adoption and also a lot of open model finetunes such as Clef.

• NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-NVFP4 by nvidia: The mid-sized model of the Nemotron family got its long awaited update. It is strong for its size and can be run at home (given you own a beefy laptop or RTX Spark). As usual, NVIDIA releases the vast majority of the data and an in-depth tech report.

• Beam-1.0-501B-A23B by reflection: The long awaited open model by Reflection, making it the strongest open American model released under a permissive license (Apache 2.0). While not the smartest open model available, Beam is very impressively token efficient (second to only GPT-6-Astra below). This is a sign of a well-developed post-training stack.