- Kimi K3.1 Pro is an open-weight, natively multimodal agentic model from Moonshot AI. It has 2.8T total parameters, 104B activated per token, and a 1,048,576-token context.
- The weights are on Hugging Face, and the API is live on platform.kimi.ai as kimi-k3.1-pro, with OpenAI and Anthropic-compatible endpoints.
- Moonshot's own benchmark table shows strong results on agentic and coding tasks and weaker ones on some others. Most numbers are self-reported, so test on your own work.
What happened
Moonshot AI has released Kimi K3.1 Pro on Hugging Face. It is post-trained from Kimi K3 with large-scale joint reinforcement learning across hundreds of downstream domains, and it is the first model in the K3.1 series. Moonshot calls it its most capable 3T-class model to date. It takes text and image input and has a 1M-token context window.
It is a Mixture-of-Experts model with 2.8T total parameters and 104B activated. It has 896 experts, 16 selected per token, and 2 shared experts. Compared with K3, Moonshot swapped the full-attention Gated MLA layers for Gated DeepSeek Sparse Attention layers to handle long context more efficiently. Weights use MXFP4 with MXFP8 activations, trained with quantization-aware training.
In Moonshot's table, Kimi K3.1 Pro (max effort) scores 65.7 on Terminal-Bench 4.0 against 66.4 for Claude Opus 5.5. It scores 92.1 on BrowseComp, 70.0 on HLE-Full with tools, and 94.2 on GPQA-Diamond (GPT-6 Astra: 96.1). It trails on some tests. Program Bench is 31.5 against 65.0 for Opus 5.5, and HLE-Diamond is 39.4 against 65.8 for GPT-6 Astra. Several benchmarks are internal, and the harnesses differ between models.
What this means for founders
- Two ways in. You can self-host the open weights or call the hosted API. The page lists vLLM, SGLang and TokenSpeed as recommended inference engines. Hardware requirements for a 2.8T model are not stated, so check the recipes before planning a self-host.
- Low switching cost for API users. Moonshot offers OpenAI and Anthropic-compatible APIs, so if your stack already uses one of those formats, trying model name kimi-k3.1-pro should be a small change.
- One integration detail can break agents. For multi-turn chats and tool calls, you must pass the full assistant message back as returned, including reasoning_content and tool_calls. Dropping them is a likely source of bugs.
- Pricing is not confirmed. The source page gives no API prices. The Hugging Face page also shows Together AI as an inference provider, but its pricing is not given. Check costs before you commit.
- Read the license. The model is under the Kimi K3.1 Pro License. Its terms are not in the source text, so read them before shipping it in a commercial product.
- Benchmarks are mixed and mostly self-reported. Many scores come from Moonshot's own runs with Claude Code or Kimi Code, and proprietary models in the table are marked with fallback behavior. Treat it as a candidate to test, not a proven winner.
How to try it today
- Open the model page at huggingface.co/moonshotai/Kimi-K3.1-Pro and read the license and the vLLM or SGLang recipe if you plan to self-host.
- For the quickest test, create an API key on platform.kimi.ai and select the model kimi-k3.1-pro through the OpenAI-compatible endpoint.
- Set reasoning_effort to "low", "high" or "max". The default is "max", so lower it for cheap or fast tasks and compare results.
- Run 5 to 10 real tasks from your product, such as coding fixes, document questions or tool-calling flows, and compare against your current model.
- For multi-turn or tool use, pass the full assistant message back unchanged, as in this pattern from the model card:
response = client.chat.completions.create(
model="kimi-k3.1-pro",
messages=messages, # include prior assistant reasoning_content and tool_calls as returned
max_tokens=4096,
reasoning_effort="max",
)Related tools on Launch Llama
Tools from the Launch Llama directory that fit with this news:
| Tool | What it does | Why it matters here |
|---|---|---|
| Kimi | For developers: Kimi K3 is an open 3T-class model with native multimodal support and 1M context window for coding and reasoning tasks. | This is the directory listing for the earlier Kimi K3, an open 3T-class model with native multimodal support and a 1M context window, which K3.1 Pro is post-trained from. |
| GPT-6 Astra | For AI product builders: GPT-6 Astra powers end-to-end applications with OpenAI's most capable model so you ship faster and scale reliably. | GPT-6 Astra is one of the closed models in Moonshot's comparison table, and it scores higher on GPQA-Diamond and HLE-Diamond. |
| LLM Reference | For engineers: LLM Reference compares 1,700+ models across 130+ providers so you pick the right LLM and lowest price in seconds. | LLM Reference compares models and prices across providers, which helps once K3.1 Pro pricing is public. Whether it lists K3.1 Pro yet is not confirmed. |
| Router by Ramp | For engineers: Router by Ramp routes every LLM request to the lowest-cost model that meets your needs so you cut inference costs and simplify API management. | Router by Ramp routes requests to the lowest-cost model that meets your needs, which is relevant if you want to mix a new model like this with others. Whether it supports K3.1 Pro is not confirmed. |
Browse all tools or list your own tool.
Llama's take
I would try this now if you build agents, coding tools or document workflows and want an open-weight option near the top of the table. The API route makes a quick test cheap in engineering time. The numbers I trust most are the ones you produce on your own tasks.
I would wait if you need a clear cost answer or plan to self-host. Pricing is not in the source, hardware needs are not stated, and the license terms are something I have not seen. The benchmarks are also uneven. Program Bench and HLE-Diamond are clear gaps against the closed models in the same table.
The one thing to watch is independent evaluation. Most of these scores come from Moonshot's own setup, so I want to see third-party results on real workloads before moving anything important.
FAQ
What is Kimi K3.1 Pro?
Kimi K3.1 Pro is an open-weight, natively multimodal agentic model from Moonshot AI. It is a Mixture-of-Experts model with 2.8T total parameters and 104B activated, and it supports a 1M-token context window. It is post-trained from Kimi K3 and is the first model in the K3.1 series.
How do I use Kimi K3.1 Pro through the API?
Use platform.kimi.ai and select the model kimi-k3.1-pro. Moonshot provides OpenAI and Anthropic-compatible APIs. Reasoning effort is set with the reasoning_effort field, which accepts low, high or max, with max as the default.
Can I run Kimi K3.1 Pro myself?
The weights are on Hugging Face, and Moonshot recommends vLLM, SGLang and TokenSpeed as inference engines, with recipes linked from the model page. The hardware needed is not stated in the source. The model is released under the Kimi K3.1 Pro License, so check its terms first.
Sources
- Kimi K3.1 Pro model weights are out (huggingface.co)
- Hacker News discussion