开源动态第 24 期：Motif-3、GLM-5.3、Hy4-preview 与开源模型许可证趋势

Avid Artifacts 的读者都知道，我们不仅长期关注模型本身，也关注它们的许可证。有一段时间，定制许可证非常流行，例如定制的 Qwen2.5 72B-Instruct 许可证：https://huggingface.co/Qwen/Qwen2.5-72B-Instruct/blob/main/LICENSE 或 Llama 许可证：https://huggingface.co/meta-llama/Llama-3.1-8B-Instruct/blob/main/LICENSE。DeepSeek 曾为 DeepSeek V3 制定了定制许可证：https://huggingface.co/deepseek-ai/DeepSeek-V3/blob/main/LICENSE-MODEL，后来 R1 将其改为 MIT：https://huggingface.co/deepseek-ai/DeepSeek-R1/blob/main/LICENSE，这导致许多中国模型开发者在 2025 年采用 MIT 或 Apache 2.0 许可证。

到 2026 年，开放模型的竞争比以往任何时候都更激烈，这带来了两个有趣的发展：西方模型开发者采用开放许可证，例如 Google：https://huggingface.co/google/gemma-4-31B-it 和 Meta：https://huggingface.co/meta-models/Muse-Glimmer-30B 都转向 Apache 2.0。而中国前沿的模型开发者则变得更加严格：Kimi K3：https://huggingface.co/moonshotai/Kimi-K3 配备了要求商业使用者签署协议的许可证，MiniMax M3：https://huggingface.co/MiniMaxAI/MiniMax-M3/blob/main/LICENSE 则要求收入超过阈值的用户签订协议，并禁止某些使用场景。

最新加入的是智谱的 GLM-5.3：https://huggingface.co/zai-org/GLM-5.3，它从 MIT（GLM-5.2 及以前版本）转为定制许可证，其中对推理和微调服务提供者包含以下条款：

如果被许可方或其任何关联公司经营模型即服务（Model as a Service）业务，并且被许可方及其关联公司在任意连续 12 个月内的总收入超过 100 亿美元（或其他等值货币），则被许可方在将软件或其衍生作品用于任何商业目的之前，必须通过 Z.AI 的安全审查。安全审查的范围和方法由 Z.AI 合理确定。

与其他同类许可相比，100亿美元的门槛非常高，但许可中并未定义“关联方”，这增加了不确定性并造成采用障碍。此外，该许可提供英文和中文版本，中文文本中使用“关联方”来表示 affiliated parties，而在中国法律中确实有定义：https://kjs.mof.gov.cn/zt/kjzzss/kuaijizhunzeshishi/200806/t20080618_46245.htm。

我们绝不是法律专家，而且这些许可的制定有明显理由。然而，我们希望强调创建此类许可所带来的问题，尤其是在存在许多有效开源和闭源替代方案的情况下。

分享：https://www.interconnects.ai/p/latest-open-artifacts-24-motif-3?utm_source=substack&utm_medium=email&utm_content=share&action=share

Motif-3：https://huggingface.co/Motif-Technologies/Motif-3 由 Motif-Technologies 提供：https://huggingface.co/Motif-Technologies：Motif 是少数几个隐藏的瑰宝之一，在模型训练上展示了创新性，即使资源远少于其他人。Motif-3 采用 MIT 许可，并且以其体量取得了令人印象深刻的成绩。根据我们在 2025 年覆盖的 Motif 2.6B：https://artifactshub.ai/Motif-Technologies/motif-2-6b 和 Motif-2-12.7B：https://artifactshub.ai/Motif-Technologies/motif-2-12-7b-instruct 发布的模型轨迹来看，其改进令人印象深刻。

dots3-note-prev：https://huggingface.co/dots-studio/dots3-note-prev 由 dots-studio 提供：https://huggingface.co/dots-studio：RedNote/Xiaohongshu，中国的 Instagram，也在更认真地进行模型训练，尽管他们并非新人：https://artifactshub.ai/rednote-hilab/dots-llm1-inst，并且早在 2025 年就已经发布了模型。dots3 还能够凭借内部工具在 IMO 2026 中获得满分。我们期待他们在不久的将来带来更多成果。

Qwen3.8-Flash-Next：https://huggingface.co/Qwen/Qwen3.8-Flash-Next 由 Qwen 提供：https://huggingface.co/Qwen：对 Qwen 模型下一版本的架构预览：125B-A6B，带有 51B n-gram 嵌入。它使用 GDN 和 Qwen 稀疏注意力。类似于 Qwen3-Next-80B-A3B-Instruct：https://artifactshub.ai/Qwen/qwen3-next-80b-a3b-instruct，我们预计相似的架构将会更受欢迎，并且在 Qwen4 发布时生态系统的集成问题将会得到解决。

GLM-5.3-Flash：https://huggingface.co/zai-org/GLM-5.3-Flash 由 zai-org 提供：https://huggingface.co/zai-org：此版本完善了我们在 2025 年撰写的中文模型手册：https://www.interconnects.ai/p/latest-open-artifacts-16-whos-building：该模型作为免费的“隐形模型”以“Ox-Alpha”的名义在 OpenRouter 和 OpenCode 上发布，这最初激发了人们尝试它的兴趣。然后，人们对其创作者和规模进行推测，声称它是 Cursor/xAI、Gemini 或开源实验室的新预训练模型，规模超过 1T。由于该模型性能相对出色，人们连续数天猜测其创作者，从而炒热了话题。这也减轻了每次（开放）模型发布伴随的极限测试指控。

Hy4-preview：https://huggingface.co/tencent/Hy4-preview 由腾讯提供：https://huggingface.co/tencent：腾讯正在成为开放模型领域的一个重要参与者，增加其旗舰模型的规模，同时推动后训练飞轮。结果，Hy4-preview 是一个能力不错的模型，目前存在过度思考的问题。不过，如果从 Hy3-preview：https://artifactshub.ai/tencent/hy3-preview 到 Hy3：https://artifactshub.ai/tencent/hy3 的发展轨迹可以参考，最终模型可能会有机会跻身开放模型的前列。

在我们的 Artifacts Hub 查看本期所有模型的更多详细信息：https://artifactshub.ai/?picks=0。

访问 artifactshub.ai：https://artifactshub.ai/?picks=0

NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16：https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16 由 nvidia 提供：https://huggingface.co/nvidia：Nemotron 的更新版本，在性能方面有提升——特别是整体速度的改进。

Ling-3.0-flash：https://huggingface.co/inclusionAI/Ling-3.0-flash 由 inclusionAI 提供：https://huggingface.co/inclusionAI：Ant Ling 经常出现在 Artifacts Log 上；他们现在已经是模型的第三次迭代，采用混合设计（KDA + Gated MLA），类似其他模型。他们还发布了一个小型 7.9B-A1.3B 版本：https://huggingface.co/inclusionAI/Ling-3.0-tiny。

Qwen3.8-2.4T-A95B：https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B 由 Qwen 提供：https://huggingface.co/Qwen：在一个相当令人惊讶的事件中，阿里巴巴也开始公开发布他们最大的 Qwen 版本。然而，它附带自定义许可，并且其性能落后于同规模的其他模型。

Avid Artifacts readers know that we have been covering not only models but also their licenses for quite some time. There was a period when custom licenses were all the rage, for example the custom Qwen2.5 72B-Instruct license：https://huggingface.co/Qwen/Qwen2.5-72B-Instruct/blob/main/LICENSE or the Llama licenses：https://huggingface.co/meta-llama/Llama-3.1-8B-Instruct/blob/main/LICENSE . DeepSeek had a custom license for DeepSeek V3：https://huggingface.co/deepseek-ai/DeepSeek-V3/blob/main/LICENSE-MODEL before R1 changed it to MIT：https://huggingface.co/deepseek-ai/DeepSeek-R1/blob/main/LICENSE , which has resulted in many (Chinese) model makers adopting MIT or Apache 2.0 licenses in 2025.

In 2026, open models are more competitive than ever, which has led to two interesting developments: Western model makers adopt open licenses, with both Google：https://huggingface.co/google/gemma-4-31B-it and Meta：https://huggingface.co/meta-models/Muse-Glimmer-30B switching to Apache 2.0. Chinese model makers at the frontier, however, are becoming more restrictive: Kimi K3：https://huggingface.co/moonshotai/Kimi-K3 comes with a license which requires commercial agreements for those who run inference or fine-tuning services, and MiniMax M3：https://huggingface.co/MiniMaxAI/MiniMax-M3/blob/main/LICENSE requires agreements above a revenue threshold and has prohibited use cases.

The newest addition is Zhipu’s GLM-5.3：https://huggingface.co/zai-org/GLM-5.3 , which switched from MIT (GLM-5.2 and earlier) to a custom license with the following clause for inference and fine-tuning providers:

If the Licensee or any of its affiliates operates a Model as a Service business, and the aggregate revenue of the Licensee and its affiliates exceeds 10 billion US dollars (or the equivalent in other currencies) in total over any consecutive 12 months, the Licensee must pass Z.AI’s security review before using the Software or its derivative works for any commercial purpose. The scope and method of the security review shall be reasonably determined by Z.AI.

While the 10 billion US dollar threshold is very high compared to other licenses of this kind, “affiliates” is not defined in the license, which adds uncertainty and creates barriers to adoption. Furthermore, the license is provided in both English and Chinese, with the Chinese text using “关联方” for affiliated parties, which does have a definition in Chinese law：https://kjs.mof.gov.cn/zt/kjzzss/kuaijizhunzeshishi/200806/t20080618_46245.htm .

We are by no means legal experts and there are obvious reasons why those licenses are created. However, we want to highlight the issues that come with creating such licenses, especially in a world with a lot of valid open and closed alternatives.

Share ：https://www.interconnects.ai/p/latest-open-artifacts-24-motif-3?utm_source=substack&utm_medium=email&utm_content=share&action=share

Motif-3 ：https://huggingface.co/Motif-Technologies/Motif-3 by Motif-Technologies ：https://huggingface.co/Motif-Technologies : Motif is one of the few hidden gems out there, showcasing innovation in their model training with very limited resources compared to others. Motif-3 comes with an MIT license and impressive scores for its size. Given the trajectory of model releases from Motif 2.6B, which we covered in 2025 ：https://artifactshub.ai/Motif-Technologies/motif-2-6b and Motif-2-12.7B ：https://artifactshub.ai/Motif-Technologies/motif-2-12-7b-instruct , the improvements are impressive.

dots3-note-prev ：https://huggingface.co/dots-studio/dots3-note-prev by dots-studio ：https://huggingface.co/dots-studio : RedNote/Xiaohongshu, the Chinese Instagram, is also getting more serious about model training, although they aren’t exactly a newcomer ：https://artifactshub.ai/rednote-hilab/dots-llm1-inst , having released models as early as 2025. dots3 was also able to win the IMO 2026 with a perfect score using an internal harness. We expect more from them in the near future.

Qwen3.8-Flash-Next ：https://huggingface.co/Qwen/Qwen3.8-Flash-Next by Qwen ：https://huggingface.co/Qwen : A preview of the next version of Qwen models in terms of architecture: 125B-A6B with 51B n-gram embeddings. It uses GDN and Qwen Sparse Attention. Similar to Qwen3-Next-80B-A3B-Instruct ：https://artifactshub.ai/Qwen/qwen3-next-80b-a3b-instruct , we expect similar architectures to become more popular and the ecosystem to fix integrations by the time Qwen4 drops.

GLM-5.3-Flash ：https://huggingface.co/zai-org/GLM-5.3-Flash by zai-org ：https://huggingface.co/zai-org : This release perfected the version of the Chinese model playbook we’ve written about in 2025 ：https://www.interconnects.ai/p/latest-open-artifacts-16-whos-building : The model got released as a free-to-use “stealth model” under the name “Ox-Alpha” on OpenRouter and OpenCode, which got people excited to try it out in the first place. They then speculated about its creator and size, alleging it is a >1T model from Cursor/xAI, Gemini or a new pre-train from open source labs. Because the model is relatively performant, people kept speculating for days about its creator, thus building up hype. It also dampens the accusations of benchmaxxing which accompany every (open) model release.

Hy4-preview ：https://huggingface.co/tencent/Hy4-preview by tencent ：https://huggingface.co/tencent : Tencent is becoming a serious player in the open model space, increasing the size of their flagship model while spinning the post-training flywheel. The result, Hy4-preview, is a competent model which currently has an issue with overthinking. However, if the trajectory from Hy3-preview ：https://artifactshub.ai/tencent/hy3-preview to Hy3 ：https://artifactshub.ai/tencent/hy3 is any indication, the final model might be a legit shot at the front ranks of open models.

View more details on all the models in this issue at our Artifacts Hub：https://artifactshub.ai/?picks=0 .

Visit artifactshub.ai ：https://artifactshub.ai/?picks=0

NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16 ：https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16 by nvidia ：https://huggingface.co/nvidia : An update to Nemotron, which comes with performance — but especially speed improvements — across the board.

Ling-3.0-flash ：https://huggingface.co/inclusionAI/Ling-3.0-flash by inclusionAI ：https://huggingface.co/inclusionAI : Ant Ling is a frequent guest at the Artifacts Log; they are now on their third iteration of models, adopting a hybrid design (KDA + Gated MLA), similar to others. They also release a small 7.9B-A1.3B ：https://huggingface.co/inclusionAI/Ling-3.0-tiny version.

Qwen3.8-2.4T-A95B ：https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B by Qwen ：https://huggingface.co/Qwen : In a rather surprising turn of events, Alibaba started to openly release their biggest versions of Qwen as well. However, it comes with a custom license and its performance is behind other models of its size.