Latest Open Artifacts: Motif-3, GLM-5.3, Hy4-preview, and Model Licenses

Open model licenses are becoming more varied

Open model coverage increasingly involves not only model capabilities, but also the licenses governing their use. Custom licenses were once particularly common, including the licenses for Qwen2.5 72B-Instruct and Llama models. DeepSeek V3 also used a custom license before DeepSeek R1 adopted MIT, after which many Chinese model developers moved toward MIT or Apache 2.0 licenses in 2025.

In 2026, open models are more competitive, and licensing trends have begun to diverge. Western model developers, including Google and Meta, have adopted Apache 2.0 for Gemma 4 31B and Muse-Glimmer-30B. Some Chinese frontier-model developers have become more restrictive. Kimi K3 requires commercial agreements for organizations that provide inference or fine-tuning services, while MiniMax M3 requires agreements above a revenue threshold and includes prohibited use cases.

Zhipu's GLM-5.3 is the latest example. GLM-5.2 and earlier versions used MIT, but GLM-5.3 uses a custom license containing the following requirement 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.

The 10 billion US dollar threshold is considerably higher than comparable thresholds in other licenses. However, the license does not define "affiliates," creating uncertainty for potential users. The license is available in English and Chinese, and the Chinese version uses the term "关联方" for affiliated parties, a term that has a definition in Chinese law.

These licensing structures have understandable motivations, but they can also create adoption barriers, particularly when users can choose among many open and closed alternatives.

Featured models

Motif-3

Motif-3 by Motif-Technologies is an example of model development with limited resources. The model uses an MIT license and posts strong results for its size. Its progress is notable when compared with the earlier Motif 2.6B and Motif-2-12.7B-Instruct releases.

dots3-note-prev

dots3-note-prev by dots-studio reflects RedNote, also known as Xiaohongshu, taking a more active role in model training. The company is not new to the field, having released models as early as 2025. dots3 reportedly achieved a perfect score at IMO 2026 with an internal harness, and further development is expected.

Qwen3.8-Flash-Next

Qwen3.8-Flash-Next provides an architectural preview of the next generation of Qwen models. It has a 125B-A6B configuration with 51B n-gram embeddings, and uses GDN and Qwen Sparse Attention. As with Qwen3-Next-80B-A3B-Instruct, similar architectures may become more common as integrations mature ahead of Qwen4.

GLM-5.3-Flash

GLM-5.3-Flash by zai-org followed a model-release pattern that attracted attention in 2025. It initially appeared as a free-to-use stealth model called "Ox-Alpha" on OpenRouter and OpenCode. Its creator and size became the subject of speculation, including claims that it might be a model larger than 1T parameters from Cursor, xAI, Gemini, or an open-source laboratory.

The model's performance sustained this speculation for several days and generated interest before its formal release. Its results also reduced some of the accusations of benchmark optimization that often accompany open model launches.

Hy4-preview

Hy4-preview by Tencent shows the company's increasing activity in open models. Tencent has increased the size of its flagship model while continuing post-training development. The current preview is capable but has a tendency to overthink. The progression from Hy3-preview to Hy3 suggests that the final model could compete among leading open models.

Other models

General-purpose models

• NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16, by NVIDIA, updates the Nemotron series with improvements in performance and, especially, speed.
• Ling-3.0-flash, by inclusionAI, is the third model generation from Ant Ling. It adopts a hybrid KDA and Gated MLA design, similar to approaches used by other developers. A smaller 7.9B-A1.3B version, Ling-3.0-tiny, is also available.
• Qwen3.8-2.4T-A95B, by Qwen, represents Alibaba's decision to release its largest Qwen variants openly. The model uses a custom license, and its performance trails other models of a similar size.