TL;DR
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
Moonshot AI has published the weights for Kimi K3.1 Pro on Hugging Face, describing it as an open-weight, multimodal model with 2.8 trillion total parameters and a 1-million-token context window. The accompanying benchmark table reports strong results on some coding and agent tasks, but it is a company-provided comparison and does not establish independent performance or practical hardware requirements.
Moonshot AI has published the Kimi K3.1 Pro model weights on Hugging Face, releasing an open-weight model the company describes as multimodal and designed for agentic tasks. Its model card specifies 2.8 trillion total parameters, 104 billion activated parameters and a context window of 1,048,576 tokens, giving developers access to the weights while leaving licensing and practical deployment details to the repository and implementation documentation.
The model card identifies Kimi K3.1 Pro as the first model in the K3.1 series, post-trained from Kimi K3 using what Moonshot calls large-scale joint reinforcement learning across hundreds of downstream domains. The stated modalities are text and images. Moonshot says the model improves coding, 3D creation, financial research, factual accuracy and other agentic capabilities; these are the company’s descriptions, not independently verified findings in the supplied material.
The listed architecture is a mixture-of-experts (MoE) design with 896 experts, 16 selected per token and two shared experts. The card reports 93 layers, including 69 KDA and 24 gated DSA attention layers, and says the model uses MXFP4 weights and MXFP8 activations with quantization-aware training. Moonshot says it replaced full-attention Gated MLA layers used in Kimi K3 with gated DeepSeek sparse attention layers to make long-context modeling more efficient.
The Hugging Face page includes a broad benchmark table comparing K3.1 Pro with named models from other companies. Results vary by task: Moonshot lists a 65.7 score on Terminal-Bench 4.0 and 51.7 on SWE Marathon v1.1, while its table gives K3.1 Pro 94.2 on GPQA-Diamond. Other listed models score higher on some of those tests. The table’s figures are not a single, directly comparable measure of overall ability: benchmark methods differ, some entries use tools or particular evaluation harnesses, and some scores are explicitly marked as internal.
Open Weights, Large Deployment Demands
Publishing weights gives researchers and developers a route to inspect, adapt and potentially run the model outside a hosted API, subject to the repository’s license and technical requirements. That can broaden experimentation with long-context and multimodal systems, including testing whether the model can handle tasks that combine text, images and tool use. Access to weights does not by itself mean that the model is easy or inexpensive to operate.
The reported scale makes deployment a central consideration. Although only 104 billion parameters are activated per token in the MoE design, the model has 2.8 trillion total parameters. The supplied model summary does not provide a complete hardware configuration, inference speed, memory requirements, or cost figures. Those details matter to teams deciding whether local use is practical, and cannot be inferred from the parameter counts alone.
The release also adds another source of results to a competitive model market where performance varies by benchmark. A high score on one test does not establish broad superiority, particularly when the publisher supplies the comparisons and evaluation conditions are not identical across all entries. Independent testing and transparent reproduction would help users judge how the reported capabilities translate to their own workloads.
Products Worth Considering
Apple MacBook Air 15-inch with M5 chip, 16GB memory, 512GB SSD, and Liquid Retina display (Midnight)
Apple MacBook Air 13-inch Laptop with M5 Chip, 16GB Memory, 512GB SSD, 12MP Center Stage Camera, Touch ID, Wi-Fi 7, Midnight
Anker SOLIX S2000 Portable Power Station, 2,010Wh, 1,500W
As an affiliate, we earn on qualifying purchases.
What Moonshot’s Model Card Reports
The release page describes Kimi K3.1 Pro as a post-trained version of Kimi K3 and the opening entry in the K3.1 series. It lists a 160,000-token vocabulary and a one-million-token context limit. Moonshot also identifies MoonViT-V2 as the vision encoder, with 401 million parameters, while the summary lists text and image as supported modalities.
The benchmark table specifies that K3.1 Pro results generally use maximum reasoning effort, temperature 1.0 and top-p 0.95 unless otherwise stated. Its footnotes describe differing evaluation setups, including use of the Claude Code harness for some coding tests and a note that certain comparisons draw from leaderboards, technical reports or other model release posts. These caveats mean figures should be read with their benchmark and evaluation conditions, rather than treated as a uniform head-to-head trial.
Licensing and Real-World Performance
The supplied source identifies a Hugging Face repository but does not include the license terms, download size, full inference instructions or required hardware. Those details determine what users may do with the weights and whether they can run the model locally. The source also does not state when the weights were posted, so the timing of the release cannot be pinned down from the material provided.
It is also unclear how the model performs under independent testing. The benchmark results come from Moonshot’s model card and include a mixture of internal evaluations and external sources, with varying harnesses and settings. The material does not provide independent replications, confidence intervals or results across a standardized set of identical runs. Claims about improvements in factual accuracy, coding and agentic ability therefore remain company-reported.
Repository Details and Independent Tests
Users considering the release can check the Hugging Face repository for its current license, files, usage guidance and any updates from Moonshot. The supplied announcement does not specify a launch schedule for related K3.1 models or describe future releases, so no additional timeline is confirmed.
The next useful evidence will be independent evaluations that document hardware, prompts, tool access and scoring methods, alongside practical reports on memory use, speed and cost. Until those details are available, the release confirms that weights have been posted, while the model’s real-world accessibility and comparative performance remain open questions.
Key Questions
What was released?
Moonshot AI posted the Kimi K3.1 Pro model weights and a model card on Hugging Face. The supplied source does not state the posting date.
How large is Kimi K3.1 Pro?
The model card lists 2.8 trillion total parameters and 104 billion activated parameters in its mixture-of-experts architecture. It also specifies a context length of 1,048,576 tokens.
Does open-weight mean anyone can use it without restrictions?
Not necessarily. The material provided describes the model as open-weight but does not include the repository’s license terms. Users should check the Hugging Face page for permitted uses and distribution conditions.
Are the benchmark results independently verified?
The supplied results appear in Moonshot’s model card and include internal tests as well as figures drawn from other sources. The material does not provide independent replication of the full comparison, so the scores should be treated as publisher-reported results.
Source: huggingface.co
