Kimi K3.1 Pro Weights: 1.56 TB, $20M License, Self-Host Guide

Moonshot AI published the full weights of Kimi K3.1 Pro on Hugging Face on November 4, a 2.8-trillion-parameter Mixture-of-Experts model with 104 billion active parameters per token and a 1-million-token context window that scores 65.7% on Terminal-Bench 4.0, behind Claude Opus 5.5's 66.4% and ahead of GPT-6 Astra's 58.2%. The 1.56 TB checkpoint requires roughly 1,874 GB of aggregate GPU VRAM, a minimum of 32 NVIDIA H100 80 GB GPUs across four nodes, putting self-hosting at approximately $190 per million output tokens versus $10 per million through the Kimi API. Moonshot's MIT-style license leaves internal use unrestricted under Section 4(a), but Section 4(b) requires model-as-a-service providers earning more than $20 million in aggregate annual revenue to sign a separate commercial agreement with Moonshot AI whose terms are not public.

Moonshot AI released the full weights of Kimi K3.1 Pro on Hugging Face https://huggingface.co/moonshotai/Kimi-K3.1-Pro on November 4. The model scores 65.7% on Terminal-Bench 4.0 — ahead of GPT-6 Astra 58.2% and within 0.7 points of Claude Opus 5.5 66.4% , at $10 per million output tokens versus Opus 5.5’s $50. That is the headline. What most coverage skips: 1.56 TB on disk, a 32-H100 minimum to run it, and a license clause that sends AI infrastructure companies to Moonshot AI’s sales team the moment their MaaS revenue clears $20M. The Benchmarks Are Real Kimi K3.1 Pro is a 2.8-trillion-parameter Mixture-of-Experts model with 104 billion active parameters per token and a true 1-million-token context window. Moonshot built it as a heavily post-trained version of Kimi K3, using an architecture of 896 routed experts 16 active per token across 93 decoder layers, with MXFP4 quantization on the routed experts and BF16 on attention and the language model head. The numbers hold up under scrutiny. On Terminal-Bench 4.0, which grades agents on realistic, resource-constrained terminal tasks, K3.1 Pro sits at 65.7% — narrowly behind Claude Opus 5.5 66.4% and clearly ahead of GPT-6 Astra 58.2% . On Moonshot’s Kimi Code Bench 3.0 long-horizon track, it scores 61.2% against Opus 5.5’s 58.5%. On HLE-Full with tools, 70.0 against Claude Fable 5.1’s 65.6. Artificial Analysis https://artificialanalysis.ai/models/kimi-k3-1-pro puts it at 55 on its Intelligence Index — the highest-scoring open-weights model tracked. The launch itself was messier. Moonshot apologized publicly after undisclosed influencer briefs pushing favorable comparisons to rivals surfaced, and rate-limiting issues burned through user quotas within minutes of launch. The technical merit stands; the conduct does not. Worth knowing before you build a workflow dependency on a vendor whose transparency record is now under scrutiny. The Hardware Math That Stops Most Teams The 1.56 TB checkpoint is 96 safetensors shards — a number that becomes concrete when you price out the cluster needed to load it. At MXFP4 precision, you need approximately 1,874 GB of aggregate GPU VRAM. That is a minimum of 32 NVIDIA H100 80 GB GPUs spread across four nodes, all connected via InfiniBand at 400–800 Gbps. At INT4, the floor drops to 19 GPUs, but throughput takes a hit. | Option | Input $/M tokens | Output $/M tokens | VRAM Required | |---|---|---|---| | Kimi K3.1 Pro API | $2.00 | $10.00 | None hosted | | Self-host MXFP4 | — | ~$190.00 | ~1,874 GB | | Claude Opus 5.5 API | $15.00 | $50.00 | None | | GPT-6 Astra API | $20.00 | $60.00 | None | Cloud cost starts at roughly $41.76 per hour for the smallest viable configuration. Run that at any realistic utilization and self-hosting comes out at approximately $190 per million output tokens — 19 times the $10 per million you pay through the Kimi API. Self-hosting only reaches API price parity if you sustain throughput above 1,500 tokens per second continuously. For most teams, that is a theoretical ceiling, not a production baseline. The License Clause Worth Reading Carefully Moonshot ships K3.1 Pro under what it calls an MIT-style license. For individual developers and enterprise teams using the model internally, that description is accurate — internal use is unrestricted under Section 4 a of the Kimi K3.1 Pro License https://huggingface.co/moonshotai/Kimi-K3.1-Pro/blob/main/LICENSE . The carve-out that matters is Section 4 b : if you operate model-as-a-service inference or fine-tuning for third parties and your aggregate annual revenue exceeds $20 million, you must enter into a separate commercial agreement with Moonshot AI before continuing. The terms of that agreement are not public — no pricing, no approval timeline, no published precedent from companies who have already crossed the threshold. A second clause requires products exceeding 100 million monthly active users or $20 million in monthly revenue to display “Kimi K3.1-Pro” prominently in their interface. If you are building an AI infrastructure product and projecting toward that revenue level, read the full license before committing K3.1 Pro to your production stack. If you are a developer running the model for internal applications, research, or fine-tuning, the license places no meaningful restrictions on your use. API vs Self-Host: When Each Makes Sense The API case is stronger than the open-weights announcement makes it sound. The Kimi API serves K3.1 Pro at $2 per million input tokens and $10 per million output, with cached input dropping to $0.20 per million. It exposes OpenAI-compatible and Anthropic-compatible endpoints, and a faster companion model — K3.1 Pro Highspeed — is available for latency-sensitive workloads at lower cost. Self-hosting makes sense in four specific scenarios: air-gapped deployments where data cannot leave your infrastructure, fine-tuning at scale where weight-level access is non-negotiable, sustained throughput above 1,500 tokens per second where hardware cost finally pencils out, and teams who already have the rack and are not provisioning new hardware for this model. AMD Instinct GPU users get Day-0 ROCm support, which broadens hardware options. Outside those scenarios, the API is the right call. The weights are on Hugging Face https://huggingface.co/moonshotai/Kimi-K3.1-Pro — 96 files, 1.56 TB, no access gate. The API is live at kimi.ai with standard authentication. vLLM and SGLang serving recipes are available for teams running inference at scale. Kimi K3.1 Pro is a genuine step forward for open-weight agentic models. Use the API unless you have a specific reason not to. If you are building MaaS infrastructure, check the license before you scale.