Companies can now run a near-frontier Chinese model on their own servers, including Huawei Ascend hardware, though weights of about 1.4TB even when compressed limit that to data-center hardware.
Moonshot AI has published the full weights of Kimi K3.1 Pro on Hugging Face, putting a 2.8 trillion-parameter model in the hands of any company with the hardware to run it. TechNode reported the release on November 5, and Shanghai’s The Paper said the weights went up in the early hours of that day, Beijing time. The model activates about 104 billion parameters per query and reads up to about 1 million tokens of context.
Moonshot had promised the release when it launched the model on October 24, with access at first through its Kimi apps and API. In its statement on the release, quoted by The Paper, Moonshot said it would keep advancing and refining open intelligence. Huawei announced inference support on the same day through vLLM Ascend and SGLang, covering its Atlas A3 and Ascend 950 systems, TechNode reported.
Moonshot also launched Kimi Hosted Agents, a managed service that runs long-lived and multi-agent tasks around the clock for internal company work or customer-facing products, The Paper reported. It gave no pricing. The release follows Moonshot’s reported confidential filing for a Hong Kong IPO of about US$3 billion, VKTR reported in September.
The weights carry a custom license and need about 1.4TB of storage
According to the model card, the model is a mixture of experts, a design that sends each piece of text to a small set of specialist sub-networks: it has 93 layers and 896 experts, of which 16 are selected for each token alongside 2 shared experts. Its context window is 1,048,576 tokens. It accepts text and images through a 401 million-parameter vision encoder and returns text. Moonshot used training that prepares the model for compression, so the weights ship in a 4-bit format (MXFP4), and lists vLLM, SGLang and TokenSpeed as supported serving software.
The scale limits who can use it. IT Home, in a report republished by Sina Tech, put the full-precision weights at about 5.6TB and still about 1.4TB in compressed 4-bit form, beyond any consumer graphics card. The Paper noted that the release falls short of open source as the Open Source Initiative defines it: training code and data stay private, and commercial use is governed by Moonshot’s own license, the Kimi K3.1 Pro License named on the model card.
K3.1 Pro is a post-trained version of Kimi K3, which Moonshot introduced in July, according to VKTR; TechNode said the new version adds reinforcement learning for coding, 3D creation, financial research and long-running agent tasks. On the Terminal-Bench 4.0 coding test, the model card shows 65.7, against 66.4 for Anthropic’s Claude Opus 5.5. IT Home reported that on Agent Arena, a ranking based on more than 17,500 real long-running agent sessions, the model placed first among open-weight models and 13th overall.
Moonshot cut back the model’s hacking skills before release
Moonshot saw the model’s vulnerability-research performance climb as post-training scaled, so it removed cyber content from the general mid-training data and cyber tasks from its reinforcement learning environments, while keeping general coding, VKTR reported when the model launched. Moonshot’s unmitigated checkpoint solved 90.5% of problems on its internal xdev-bench-hard exploit test, against 81% for the released Kimi K3.1 Pro and 33.3% for Kimi K3. Moonshot said it aims "not to release open weights that would extend the absolute frontier of cyber capability."
Open weights also mean those choices can be undone. At least one third-party repository on Hugging Face already lists K3.1 Pro as its base model: a "de-risked, cyber-hardened" derivative that says it subtracts the model’s refusal direction from selected layers and fine-tunes it on a cybersecurity corpus. Its card says its own weights have not yet been uploaded and that no benchmark or safety results have been published.
Huawei’s Ascend support gives buyers a domestic route
Same-day support on Ascend means Chinese companies can deploy the model on domestic accelerators from the day the weights appeared. The Paper said the model can be deployed and validated directly on Atlas-series hardware. That fits a wider move among Chinese labs: DeepSeek is reported to be planning to pre-train V5 on Huawei’s Ascend chips.
The release adds a large open-weight competitor for Alibaba’s Qwen models. The next measures to watch are download counts for the official repository, which already shows 11,658 for the past month, and whether cloud providers in China begin offering K3.1 Pro on Ascend capacity.