Moonshot ships Kimi K3.1 Pro API, open weights due November 4

Moonshot AI put Kimi K3.1 Pro on the Kimi API this morning with a commitment that matters more than the launch itself: full model weights by November 4. The 2.8-trillion-parameter mixture-of-experts model scores 55 on Artificial Analysis' independent Intelligence Index, third among tracked models, at $2 per million input tokens.

Key facts

• Kimi K3.1 Pro scores 55 on the Artificial Analysis Intelligence Index, ahead of OpenAI's GPT-6 Astra and Google's Gemini 4 Argon (53 each) and behind Anthropic's Claude Opus 5.5 (58)
• Moonshot commits to releasing the full model weights by November 4, 2026
• Published API rates list $2.00 per million input tokens and $10.00 per million output tokens, with cached input at $0.20
• Moonshot reports 65.7% on Terminal-Bench 4.0, ahead of GPT-6 Astra (58.2%) and just behind Claude Opus 5.5 (66.4%) on that vendor-run benchmark
• The model carries a 1-million-token context window, and a Highspeed variant runs 2-3× faster, according to Moonshot

What happened

Kimi K3.1 Pro is a post-trained version of Kimi K3, the 2.8-trillion-parameter mixture-of-experts model Moonshot AI released in July, retrained with substantially more post-training compute. It is live now on Kimi.ai, Kimi Code, Kimi Work and the Kimi API, and Moonshot's announcement states the plan plainly: "Full model weights will be released by Nov 4." Developer Simon Willison confirmed the API was serving the model on launch day under the name kimi-k3.1-pro through Moonshot's OpenAI-compatible endpoint.

Two variants shipped. Moonshot says K3.1 Pro Highspeed delivers 2-3× the speed of the standard model; we have not measured that ourselves. Pricing on the standard model comes to $2.00 per million input tokens and $10.00 per million output tokens, with cached input at $0.20, roughly a third below the $3/$15 Kimi K3 launched at in July.

Moonshot's benchmark story centers on agentic coding. The company reports 65.7% on Terminal-Bench 4.0, ahead of GPT-6 Astra's 58.2% and narrowly behind Claude Opus 5.5's 66.4%, plus a claimed hallucination-rate drop from 23.4% to 8.3% on its internal stress-test benchmark. Those numbers come from Moonshot's own harness. The independent check is Artificial Analysis, which scores K3.1 Pro at 55 on its Intelligence Index, ahead of GPT-6 Astra and Gemini 4 Argon (53 each) and behind Claude Opus 5.5 (58). The Decoder notes that if the weights ship as planned, K3.1 Pro would become the highest-scoring open-weights model available.

One disclosure in the announcement deserves builder attention. Moonshot says it identified and removed cyber-related data from both the mid-training corpus and the reinforcement-learning environment mix ahead of the open-weight release, after observing vulnerability-research capability rising with post-training scale. That means the downloadable November weights will have been trained under a different data mix than the model behind today's API scores, at least on security-adjacent tasks. It is a disclosed change, not an independent measurement, but it is the first time a major lab has documented pre-release data removal of this kind for an open-weight drop.

Why it matters

For builders deciding where to run models, today's launch is an API product and November 4 is a decision point. If the weights land with scores intact, the open-weights frontier moves to roughly three points behind the strongest closed model on the Artificial Analysis index. That compresses the gap that has kept capability-sensitive workloads on closed APIs, and it lands three days before Mistral's promised Large 4 weight release, which we flagged in this week's roundup.

The serving math is the catch. K3.1 Pro inherits Kimi K3's shape: 2.8 trillion total parameters with about 104 billion active per token across 896 experts. Self-hosting a model at that total size means multi-GPU or multi-node hardware with expert-parallel serving, not a single 24GB workstation. The open-weight releases of the past two weeks have made the same point at 501B and 816B parameters. So the realistic plan for most teams is to ride the $2/$10 API now, benchmark the weights on November 4, and price self-hosting against that bill. Our guide to running Kimi models through the API covers the endpoint setup in the meantime.

The license terms ship with the weights, and they are worth reading before anything commercial is planned. October's two big US open-weight drops took opposite approaches, with Reflection's Beam shipping clean Apache 2.0 while Meta's Muse Spark Open carries attribution strings in its custom license, as we covered when Muse Spark's memory-serving stack launched. Which bar Moonshot sets will shape whether K3.1 Pro shows up in managed routers or stays a direct-API product.

Background

Kimi K3 arrived in July as Moonshot's first 3T-class mixture-of-experts model, with 2.8 trillion total parameters, about 104 billion active per token, native vision, and a 1-million-token context window. K3.1 Pro is the first release in the K3.1 series, and Moonshot attributes its gains to scaled reinforcement learning across hundreds of downstream domains rather than architecture changes.

The launch extends an October that has already been unusually dense for open weights. Reflection AI's Beam weights download and serving guide shipped a 501B Apache 2.0 model last week, DeepSeek's V4.1 Flash update refreshed the efficiency tier, and NVIDIA's Nemotron 3.5 pair split the difference between 1M-context scale and 24GB local fit. Artificial Analysis has become the neutral scoreboard for all of it, which is why its 55 for K3.1 Pro carries more weight with us than any vendor table.

Moonshot's cyber-data removal is also a precedent worth tracking. Labs have disclosed capability controls before, but publishing the removal itself, with the training-stage detail, sets a template other labs may follow when frontier-adjacent weights go public. How much capability the removal actually costs will be measurable once the community benchmarks the November weights against today's API results.

What's next

November 4 is the date. Three checks on the day: the license text and its commercial terms, the hardware fit for a 2.8T-parameter MoE (quantization and expert-parallel support), and whether Artificial Analysis re-scores the open weights at the same 55. Mistral Large 4's open weights are committed for October 27 per Reuters, so the week ends with two frontier-adjacent weight drops to evaluate. Watch also whether Highspeed's premium pricing reaches the public API docs, since the launch post describes the speed but not the exact rate.