Kimi K3.1 Pro weights are out: the 2.8T model card answers what the launch post left open

Moonshot AI released Kimi K3.1 Pro on October 23 and promised the full model weights by November 4. The weights kept that date: the moonshotai/Kimi-K3.1-Pro model page was published on Hugging Face on November 4, and the model card that comes with it is more concrete than the launch post was. It confirms 2.8 trillion total parameters with 104 billion active per token, 896 experts, a 1,048,576-token context window, native image input, and a custom license. It also documents one API behavior that will silently degrade your agent if you miss it. The vendor benchmark table is more honest than most: K3.1 Pro leads on several agentic and coding benchmarks and loses clearly on others. If you run coding agents or long-context pipelines, this is worth an evaluation slot. If you hoped to self-host it on a single node, the size ends that thought early.

What the model card confirms

• Architecture: MoE with 2.8T total and 104B activated parameters, 93 layers, 896 experts with 16 selected per token plus 2 shared experts, and a 160K vocabulary.
• Attention: a hybrid of KDA layers and Gated DeepSeek Sparse Attention (Gated DSA) layers, which replace the full-attention Gated MLA layers from Kimi K3 for more efficient long-context modeling. Context length is 1,048,576 tokens.
• Vision: a MoonViT-V2 encoder with 401M parameters, so the model takes text and images natively.
• Quantization: MXFP4 weights and MXFP8 activations, trained quantization-aware rather than quantized after the fact.
• Access: open weights on Hugging Face under a custom "Kimi K3.1 Pro License" (not Apache or MIT; read it before commercial use), plus a hosted API as model kimi-k3.1-pro on platform.kimi.ai with OpenAI- and Anthropic-compatible protocols. Recommended self-hosting engines: vLLM, SGLang and TokenSpeed, each with published recipes. The launch post also mentions a faster Highspeed variant; I did not find Highspeed weights in what I read, so treat it as API-only for now.

The benchmark picture, with warnings attached

All numbers below are Moonshot's own table, run at reasoning effort max, so treat them as vendor claims. Within that table, the picture is genuinely mixed, which is more credible than a clean sweep:

• Leads: SWE Marathon v1.1 at 51.7 (Claude Opus 5.5 at 46.3, GPT-6 Astra at 42.5), HLE-Full with tools at 70.0 (Opus 67.7, Astra 57.2), BrowseComp at 92.1, MMMU-Pro with python at 90.0.
• Competitive: Terminal-Bench 4.0 at 65.7, just behind Opus 5.5's 66.4; DeepSWE at 75.5, behind DeepSeek-V4.1-Flash's 77.0.
• Clear losses: Program Bench at 31.5 against Opus 5.5's 65.0, and Terminal-Bench-Science at 38.1 against GPT-6 Astra's 68.1.

Two rows are marked internal (Kimi Code Bench 2.0 and 3.0), so they carry even less weight than the rest. The honest reading: K3.1 Pro is in the front pack for agentic coding and long-horizon work, not uniformly ahead of it.

What I would check

My untested suggestions before routing real workloads:

• Run your own coding-agent evals against both the API and a self-hosted endpoint; the vendor table shows task-family sensitivity is large (a 33-point swing between Program Bench and SWE Marathon).
• Budget for max reasoning effort: every published score uses it, and the default is "max" with "low" and "high" available. Cheaper settings will not match the table.
• Handle the multi-turn contract carefully: the API requires the complete assistant message passed back as-is, including reasoning_content and tool_calls, not just content. Clients that strip reasoning between turns will degrade silently.
• Read the custom license before anything commercial; it is its own document, not a standard OSS license.
• If self-hosting, plan around MXFP4 weights from day one; 2.8T parameters even at 4-bit is a multi-node serving problem.

What I would not conclude

Do not conclude this beats the frontier on coding in general; the vendor's own table shows two stark losses. Do not assume the 1M context is cheap or fast at full length; the card gives no latency or cost curve at long context. Do not treat the Hugging Face publication as the moment the model became usable to you: engine recipes and API behavior are what matter, and those are days old at most. And do not read the cybersecurity benchmark footnotes as evidence of real-world exploit skill; those are capability ceilings under artificial budgets, reported by the vendor.

Frequently asked questions

Are the Kimi K3.1 Pro weights really available now?

Yes. The moonshotai/Kimi-K3.1-Pro repository on Hugging Face was published on November 4, 2026, matching the date the October 23 launch post promised. The license is a custom "Kimi K3.1 Pro License", not a standard OSI license, so read it before building on the weights.

Can I run Kimi K3.1 Pro locally?

Not on a workstation. 2.8T total parameters with MXFP4 weights is a multi-GPU, multi-node serving job. Moonshot recommends vLLM, SGLang or TokenSpeed with published recipes. Everyone else should use the hosted API.

How do I call the API?

On platform.kimi.ai, select the model kimi-k3.1-pro. The API is compatible with both OpenAI and Anthropic client protocols, so existing SDK code mostly carries over.

What is the one API behavior most likely to break my agent?

Multi-turn state. You must pass the complete assistant message back in messages, including reasoning_content and tool_calls, not just the text content. Dropping the reasoning between turns is the silent failure mode.

How does it compare to Claude Opus 5.5 or GPT-6 Astra?

On Moonshot's own table: ahead on SWE Marathon, HLE-Full with tools, BrowseComp and MMMU-Pro; roughly tied on Terminal-Bench 4.0 and DeepSWE; clearly behind on Program Bench and Terminal-Bench-Science. All vendor-reported at max reasoning effort.

Is the 1M token context real?

The card lists a context length of 1,048,576 tokens and the architecture changes (Gated DSA replacing full-attention MLA layers) exist specifically to make long context efficient. The card does not publish long-context pricing or latency, so verify cost on your workload before committing.

What changed from Kimi K3?

K3.1 Pro is post-trained from K3 with large-scale joint reinforcement learning across downstream domains, swaps the full-attention Gated MLA layers for Gated DSA layers, and adds the MoonViT-V2 vision stack. Moonshot calls it its most capable 3T-class model to date.

Sources

• moonshotai/Kimi-K3.1-Pro model card | Hugging Face - read November 8, 2026. Architecture table, evaluation table and footnotes, deployment engines, the reasoning_effort and multi-turn message requirements, and the license statement.
• Kimi K3.1 Pro launch post | kimi.ai - October 23, 2026, read November 8, 2026. The release announcement, the Highspeed variant, and the promise that full weights would be released by November 4, 2026.
• Moonshot AI organization page | Hugging Face - read November 8, 2026. Publication timeline: the K3.1 Pro model page was published November 4, 2026 and updated since.

Not read: the K3.1 Pro License text itself, the vLLM/SGLang/TokenSpeed recipes, the Kimi K3.1 Pro Quickstart, and any independent evaluations. All benchmark figures above are vendor-reported. I have not run the model.