Kimi K3.1 Pro Open Weights: 2.8T MoE, 1M Context, $2 API Price

November 5, 2026 | AI | By Mahesh

Moonshot AI has published Kimi K3.1 Pro as open weights on Hugging Face, giving developers a downloadable multimodal model with 2.8 trillion total parameters and a 1,048,576-token context window.[2] The Chinese lab first released the model through its apps and API on October 23, 2026, and committed to posting the weights by November 4.[1] The release targets teams that want coding and agent performance close to the top closed models while keeping inference on hardware they control.

What is new in Kimi K3.1 Pro

Kimi K3.1 Pro is a post-trained version of Kimi K3 that uses much larger reinforcement learning compute, a one-million-token context window and a focus on coding, 3D creation and financial research.[1] It is the first model in Moonshot's K3.1 series. A companion variant, Kimi K3.1 Pro Highspeed, serves the same model two to three times faster for interactive use.[1]

"our most capable 3T-class open-weight model to date"

Moonshot AI, Kimi K3.1 Pro Official Announcement [1]

The gains over K3 are largest in agentic work. On Kimi Code Bench 3.0 Long-horizon, an internal Moonshot test, K3.1 Pro scores 61.2% against 55.95% for K3.[1] On OfficeQA Pro v2, which tests retrieval and numerical reasoning over financial documents, it scores 64.4, ahead of Claude Opus 5.5 at 62.2 in Moonshot's comparison table.[2] Moonshot also reports that the claim-level hallucination rate on its internal Kimi Hallu Bench Hard fell from 23.4% for K3 to 8.3%. The company says that test is a deliberate stress set and does not represent everyday error rates.[1]

The model also adds stronger 3D and CAD ability. On BenchCAD with tools, which asks a model to rebuild a 3D shape from multi-view renderings through CAD code, K3.1 Pro scores 92.7% against 96.2% for Claude Opus 5.5 and 95.9% for GPT-6 Astra.[1] A new 3D Make feature on kimi.ai lets users preview and edit generated models and export them for 3D printing.[1]

Under the hood: architecture and specifications

Kimi K3.1 Pro is a mixture-of-experts model with 2.8 trillion total parameters, 104 billion activated per token and 93 layers.[2] Each token is routed to 16 of 896 experts, plus 2 shared experts. The attention stack mixes 69 Kimi Delta Attention layers with 24 Gated DeepSeek Sparse Attention layers, and Moonshot says the sparse layers replace the full-attention Gated MLA layers used in K3 to make long-context work cheaper.[2]

The model accepts text and image input through a 401 million parameter MoonViT-V2 vision encoder and returns text.[2] Weights ship with MXFP4 quantization and MXFP8 activations, trained with quantization-aware training, so the 4-bit format is the intended serving format.[2] Reasoning is always on, with a reasoning_effort field that accepts low, high and max, and max is the default.[2] For multi-turn and tool use, the API requires the full assistant message, including reasoning_content and tool_calls, to be passed back unchanged.[2]

Benchmarks and performance: reported results against marketing claims

Moonshot reports 65.7% on Terminal-Bench 4.0, behind Claude Opus 5.5 at 66.4% and ahead of GPT-6 Astra at 58.2% and Claude Fable 5.1 at 57.9%.[1] Every K3.1 Pro figure comes from Moonshot's own runs at maximum reasoning effort with the Claude Code harness. Peer scores are taken from vendor reports and public leaderboards, so the rows below compare numbers produced under different conditions.[2]

Source: Moonshot AI model card, maximum reasoning effort, with Claude-family and GPT-6 Astra scores taken from vendor reports and leaderboards.[2]

The table shows a model that leads or ties on several agentic and document tasks and trails by a wide margin on others. K3.1 Pro tops DeepSWE and SWE Marathon in this comparison, yet scores 31.5 on ProgramBench against 65.0 for Claude Opus 5.5 and 39.4 on HLE-Diamond against 65.8 for GPT-6 Astra.[2]

Independent numbers are thinner. AiCybr, citing Artificial Analysis, reports an Intelligence Index score of 55, fourth among 215 models on that site's page on October 24, with a measured Terminal-Bench 4.0 result of 67%, about 73.5 output tokens per second and 3.20 seconds to first token.[5][6] Those figures were measured on the hosted API and can shift as the evaluator adds samples.

Pricing, availability and how to access Kimi K3.1 Pro

The hosted Kimi K3.1 Pro costs $2.00 per million input tokens and $10.00 per million output tokens, with cache hits at $0.20 per million.[4][6] Kimi K3.1 Pro Highspeed costs $4.00 and $20.00. Kimi K3 is listed at $3.00 and $15.00, so the new standard tier is one third cheaper than its predecessor.[6]

Developers can call the model through an OpenAI-compatible endpoint with the IDs kimi-k3.1-pro and kimi-k3.1-pro-highspeed.[4] The model also runs inside the Kimi apps, in Kimi Code through the /model command and in the Kimi Work desktop app, version 3.1.0 or later, on Windows and Apple silicon Macs.[1] Enterprise administrators must switch it on, because access is off by default, and Zero Data Retention is available to API customers.[1] For self-hosting, the model card recommends vLLM, SGLang and TokenSpeed, with a published vLLM recipe.[2][7]

The license is custom and permissive for most users. It allows use, modification, fine-tuning and redistribution, with two conditions for large operators.[3] A company that runs a Model as a Service business and earns more than $20 million over any 12 months needs a separate agreement with Moonshot before commercial use. A product with more than 100 million monthly active users or more than $20 million in monthly revenue must show "Kimi K3.1-Pro" in its interface. Internal use and access through Moonshot's official products or certified inference partners are exempt from both conditions.[3]

Limitations, trade-offs and competitive alternatives

Running Kimi K3.1 Pro yourself needs a cluster. The 104 billion active figure describes compute per token, and the checkpoint still holds all 2.8 trillion parameters.[2] At 4 bits per weight that is roughly 1.4 TB of raw weights by simple arithmetic, before key-value cache and runtime overhead, against 320 GB for four 80 GB accelerators.[6] Most teams will reach the model through the API or a certified provider.

Moonshot also trimmed offensive security skill before release. It removed cyber-related data from mid-training and from its reinforcement learning environments, and its internal xdev-bench-hard pass@3 score fell from 90.5% to 81.0% as a result.[1] The company says the final model still trails frontier proprietary systems that have been public for at least three months. On ExploitBench, K3.1 Pro scores 74.7 against 83.0 for Claude Fable 5.1, 91.0 for Claude Opus 5.5 and 100 for GPT-6 Astra.[2] Several headline tests are internal, including Kimi Code Bench and Kimi Hallu Bench, so teams should rerun their own tasks before trusting them. Artificial Analysis data cited by AiCybr also shows heavy token use, about 150 million output tokens across its index, which raises real task cost above the list price.[6]

Open-weight alternatives at this scale include Tencent's 770 billion parameter model covered in Tencent Hy4: 770B MoE, 1M Context, 6 Yuan per Million Tokens and the 1.05 trillion parameter preview in Mistral Large 4 Preview: 1.05T Open-Weight MoE, Specs and Pricing. Teams that want the highest scores on ProgramBench or HLE-Diamond will still find them on the closed models in the table above.

Sources

1. Moonshot AI. "Introducing Kimi K3.1 Pro," October 23, 2026. kimi.ai
2. Moonshot AI. "Kimi-K3.1-Pro model card and evaluation results," Hugging Face. huggingface.co
3. Moonshot AI. "Kimi K3.1-Pro License," Hugging Face. huggingface.co
4. Moonshot AI. "Kimi API Platform and K3.1 Pro quickstart," pricing and model IDs. platform.kimi.ai and quickstart guide
5. Artificial Analysis. "Kimi K3.1 Pro model page," independent evaluation. artificialanalysis.ai
6. AiCybr Editorial. "Kimi K3.1 Pro: API Pricing, Benchmarks, Open-Weight Release and Self-Hosting Requirements," October 24, 2026. aicybr.com
7. vLLM Project. "Kimi-K3.1-Pro serving recipe." recipes.vllm.ai

Read More on Depth Grid

• Tencent Hy4: 770B MoE, 1M Context, 6 Yuan per Million Tokens
• Mistral Large 4 Preview: 1.05T Open-Weight MoE, Specs and Pricing
• Grok Voice 3.0 Launch: $0.08 per Minute, 86.9 Index, Specs
• AI Native Startup Playbook: Six Moves From Idea to Series A

Article by Mahesh | Depth Grid