This observation covers verified technical changes formally made public from approximately 7:54 am on 4 November to 7:54 am on 5 November 2026. It includes two items.
Kimi K3.1 Pro model weights are delivered
Moonshot AI published the Kimi K3.1 Pro weights on Hugging Face on 4 November. The model had been available through Kimi products, Kimi Code and the API since 23 October, while the complete weights were still planned. The new fact is that downloadable files, configuration, tokenizer, custom model code and the licence are now in a public repository. The repository is about 1.56 TB and contains 96 Safetensors weight shards. The model card specifies 2.8 trillion total parameters, 104 billion parameters activated per token, and a quantisation-aware format using MXFP4 weights and MXFP8 activations. Moonshot provides deployment paths for vLLM, SGLang and TokenSpeed.
The weights use the custom Kimi K3.1 Pro License rather than Apache 2.0 or MIT. It permits use, modification, distribution, fine-tuning and derivative works, but a Model-as-a-Service operator whose combined revenue with affiliates exceeds US$20 million over any consecutive 12 months must enter a separate agreement with Moonshot AI before commercial use. Commercial products with more than 100 million monthly active users or more than US$20 million in monthly revenue must also display the model name prominently. Capability and benchmark results in the model card remain largely developer evaluations, and the weight release had not yet produced a complete independent reproduction by the cut-off.
Tencent Hy4 enters cloud services while the weights remain pending
Tencent released the production version of Hy4 on 4 November, and Tencent Cloud added it to ADP and TokenHub on the same day. Compared with the Hy4 Preview released on 28 August, the production version expands pre-training and post-training, adds visual understanding and computer use, and strengthens long-document comprehension, long-horizon tasks and deep research. Tencent Cloud ADP states that the service supports about 1.05 million tokens of context and up to 65,536 output tokens per request, with exact limits depending on the access channel. The model can combine graphical interfaces, command-line operations and tool calls for multi-step work.
Hy4 is currently available as a service, while the production weights have not been delivered. Tencent published results for finance, game development, scientific research, search hallucination and safety tests, but these figures come from Tencent and partner testing and have not been fully reproduced independently. Reuters verified Tencent's account that cybersecurity training material was restricted to defensive uses such as identifying and fixing vulnerabilities, with material on building attack tools, breaking into systems and penetration testing excluded; the risk assessment itself remains based on Tencent's internal security teams. Tencent says the production weights and a safety testing report will be released later, so the current status is not a completed open-weight delivery.
Sources
Moonshot AI, Introducing Kimi K3.1 Pro, 23 October 2026, https://www.kimi.com/en/blog/kimi-k3-1-pro
Hugging Face, moonshotai/Kimi-K3.1-Pro file repository, 4 November 2026, https://huggingface.co/moonshotai/Kimi-K3.1-Pro/tree/main
Moonshot AI, Kimi K3.1 Pro License, 4 November 2026, https://huggingface.co/moonshotai/Kimi-K3.1-Pro/blob/main/LICENSE
Tencent Cloud ADP, Tencent Hunyuan Hy4 production release on Tencent Cloud ADP, 4 November 2026, https://adp.tencent.com/zh/blog/tencent-cloud-adp-hunyuan-hy4-release
Tencent Cloud, TokenHub product updates, 4 November 2026, https://cloud.tencent.com/document/product/1823/130675
Reuters, Tencent rolls out AI model trained with limited cybersecurity data amid hacking fears, 4 November 2026, https://www.reuters.com/legal/litigation/tencent-rolls-out-ai-model-trained-with-limited-cybersecurity-data-amid-hacking-2026-11-04/
Discover more from Geoffrey Chen
Subscribe to get the latest posts sent to your email.