What happened
The Hugging Face repository identifies Kimi K3.1 Pro as an open-weight, natively multimodal agentic model and Moonshot AI’s most capable 3T-class model to date. The model card says it is the first model in the K3.1 series and was post-trained from Kimi K3 with large-scale joint reinforcement learning across hundreds of downstream domains. Moonshot says it has a 1-million-token context window and improvements in coding, 3D creation, financial research, factual accuracy and other agentic capabilities.
The card describes a mixture-of-experts architecture with 2.8 trillion total parameters and 104 billion activated parameters. It lists 93 layers, 896 experts, 16 selected experts per token, and two shared experts. The context length is 1,048,576 tokens. Supported modalities are text and image, with a 401-million-parameter MoonViT-V2 vision encoder. Weights use MXFP4 and activations MXFP8, which the card describes as quantization-aware training. Moonshot also says it replaced full-attention Gated MLA layers used in Kimi K3 with Gated DeepSeek Sparse Attention layers for more efficient long-context modeling.
The repository includes starting points for Transformers, vLLM, SGLang and Docker Model Runner. Its vLLM and SGLang examples launch a local server and call an OpenAI-compatible chat-completions endpoint; the example request includes an image URL as well as text. The Transformers instructions show loading the model with trust_remote_code=True. These examples establish documented routes to try the weights, but the supplied material does not give hardware requirements, expected throughput or a deployment cost estimate.
The model card reports results across coding, agentic, reasoning, knowledge and multimodal benchmarks, alongside selected competitors. It says that, unless otherwise specified, Kimi K3.1 Pro results use maximum reasoning effort, temperature 1.0 and top-p 0.95. The table includes Terminal-Bench 4.0 at 65.7, SWE Marathon v1.1 at 51.7 and BrowseComp at 92.1. Moonshot cautions that benchmark metrics differ and are not directly comparable across score types. The card names a Kimi K3.1 Pro license, but the supplied excerpt does not provide its terms.
What the archive adds
This is a new fact on a known story: the model card now says the weights are open and provides deployment instructions. Earlier coverage reported Moonshot’s plan to release them by Nov. 4: Moonshot AI has added Kimi K3.1 Pro to its Kimi Research Blog. A related note covered the stated cyber-data training measure before release: Moonshot says Kimi K3.1 Pro was trained with cyber-related data removed to cap offensive capability before open-weight release.
Trend signal
This supports the pattern Harness architecture outlives model choice: the card makes the model available through multiple serving and application paths, including vLLM, SGLang and Transformers. That gives builders options for integrating it into different harnesses. The source does not show how these integrations behave in production, so it is an access signal rather than evidence of operational performance.
Why now
The archive says Moonshot had promised open weights by Nov. 4; the Hugging Face repository now lists them. The timing marks delivery of a previously announced access path, not an unexplained change in the model’s capabilities.
Why it matters
Open weights give teams a new option to evaluate the model in their own deployment environment and compare that experience with using Moonshot’s API. The repository’s examples reduce the first-step uncertainty for teams already using Transformers, vLLM or SGLang: they can follow a documented route to load or serve the model and test text-and-image requests.
That access does not establish that self-hosting is cheaper, faster or operationally preferable. A 2.8-trillion-parameter model is a substantial deployment proposition, but this source does not state the hardware configuration required or the costs of running it. Teams should treat Moonshot’s benchmark table as a vendor-reported starting point, not a substitute for task-specific evaluation. The useful decision is whether the weights and documented interfaces let a team test its own workload and deployment constraints.
What happens next
- Teams using vLLM or SGLang will begin testing Kimi K3.1 Pro against internal text-and-image workloads. (3 months, 65%). Watch for: published deployment configurations, latency figures, and reproducible task evaluations.
Limits and uncertainty
The supplied source is Moonshot’s Hugging Face model card, so its capability descriptions and benchmark results are vendor claims. The card reports scores and settings but does not establish independent reproduction, performance on a reader’s workload, or whether differences in evaluation setup affect comparisons. Some benchmark names are marked internal, and the card itself warns that benchmark metrics are not directly comparable across metric types.
The supplied excerpt does not state the license terms, total download size, hardware requirements, inference speed, memory needs, hosted pricing or availability of particular inference providers. It also does not document safety evaluations beyond its benchmark sections. The presence of launch instructions is evidence of a documented deployment route, not proof that every configuration will work without adaptation.
Skill to build now
Take one representative text-and-image task from your workload and run it through a small evaluation harness using a deployment route your team already supports, such as vLLM or SGLang. Record the exact model revision, hardware, latency, failure cases and output quality. Keep vendor benchmark scores separate from your own results, and check the repository’s license terms before building a service around the weights.
Reader takeaway
Kimi K3.1 Pro’s weights are now available on Hugging Face, with examples for several familiar deployment tools. That makes a hands-on evaluation possible; it does not, on its own, answer whether the model is affordable or suitable for a particular production workload.