Kimi K3.1 Pro

Kimi K3.1 Pro is Kimi's post-trained K3 flagship, callable today, with the weights still set for November 4.

You can call it now. You cannot download it yet.

The tech blog dated the rollout to the day it went up, and Kimi Code's notes date that day October 23, 2026. This page did not call the model.

• K3.1 Pro and K3.1 Pro Highspeed are on Kimi.ai, the iOS, Android, and HarmonyOS apps, Kimi Work 3.1.0 or later for Windows and Apple silicon Macs, Kimi Code, and the API. Model ids are kimi-k3.1-pro and kimi-k3.1-pro-highspeed. The API base is api.moonshot.ai.
• The blog calls it a 3T-class open-weight model and a post-trained K3, with a 1 million token window. It does not print a parameter split. The previous K3, on the API homepage, is the one labeled 2.8T.
• Full weights are scheduled for November 4, 2026. They are not a download today.
• Enterprise access is off unless an administrator turns it on. Kimi Code puts K3.1 Pro on Plus, also called Moderato, and above. The 1 million token window, and Highspeed, need Pro, also called Allegretto, or above.
• Highspeed is the same model run faster. The blog says 2 to 3 times. The API homepage says about 3 times, and "same intelligence."

The API bill is flat, and thinking stays on

On October 24, 2026, the API homepage priced a million tokens like this. A flagship call also needs a top-up of at least 1 dollar before the model unlocks.

• K3.1 Pro: 2 dollars of input, 10 dollars of output, 2 dollars to write the cache, 0.20 dollars for a cache hit.
• Highspeed: 4, 20, and 4 dollars for those first three, and the same 0.20 dollar cache hit.
• The homepage prints one cache-write price. The docs say a write is billed on a 5-minute or 1-hour timer. Those two timers are not given two different prices on the homepage. There is no extra charge for a long context. The previous K3 on the same page is 3, 15, 3, and 0.30 dollars.
• Thinking cannot be switched off. Effort is low, high, or max, and max is the default. Kimi Code returns an error if thinking is disabled. In the API, temperature, top-p, and the repetition penalties are fixed. Leave them out. The reply cap starts at 131,072 tokens and can be raised to 1,048,576.
• A picture has to be sent as data in the request, or as an uploaded file id. A public image link is not accepted. The docs say web search is being rebuilt and is not for production use yet.
• Do not switch an existing K3-series, K2.8 Preview, or K2.7 Code Highspeed session over to K3.1. Kimi Code says that can make the model behave oddly and waste the cache. Start a new session.

Their scores, and the harness attached to them

The numbers below are from the tech blog. The footnotes say K3.1 Pro was run at max effort, temperature 1.0, and top-p 0.95, unless a line says otherwise. They are not a test run for this listing.

• Terminal-Bench 4.0: 65.7 percent, using the Claude Code harness. They place that above GPT-6 Astra at 58.2 and just under Claude Opus 5.5 at 66.4. Those two comparison figures are the highest number they took from a blog post or the public leaderboard, not a rerun in the same harness.
• Their own Kimi Code Bench 3.0 Long-horizon: 61.2 percent, against K3 at 55.95. That bench is internal.
• BenchCAD, with tools: 92.7 percent. Opus 5.5 is 96.2 and GPT-6 Astra is 95.9 in the reports they cite. Their own open-model runs used a 1,000-case slice of a 17,900-case set. The proprietary numbers are not from that slice.
• OfficeQA Pro v2: 64.4, which they call the best result they have. The agent gets the financial PDFs as pictures, not as extracted text. If another model looked the answer up online, they scored that task zero.
• Against K3, on their own finance tests: modeling up 7.2 points, research up 11.6 points, and a simulated business about 4.2 times the earlier score.
• On their hard hallucination stress test, claim-level errors go from 23.4 percent on K3 to 8.3 percent. Answers that still contain at least one factual error are 42.4 percent of K3's rate. They say the prompts were built to trip a model, and the rates are not an everyday error rate. The API docs page describes a different check, with different percentages, and still has blank score slots in another paragraph. The figures here are the blog's.

What they held back for the weight release

They say cyber-related material was taken out of mid-training and out of the reinforcement-learning mix before the open-weight checkpoint. On their internal xdev-bench-hard, the unfiltered checkpoint solves 90.5 percent of tasks at pass@3, and the cleaned K3.1 Pro solves 81.0 percent. They also say that cleaned score is still behind proprietary frontier models that have been public for at least three months. The weights that would let someone run this locally are the November 4 release, not a file on the blog today.

One internal coding story on the blog: a walking simulator, built on an open anatomical model, went from falling over to a continuous walk of 11.2 seconds after 167 missions. That is their account of their own run. 3D Make, for previewing and exporting those models, is live on kimi.ai and in the app.

Use it at kimi.ai, in Kimi Code with /model, or on the API platform. The picture on this page is the blog's share card, mountains and a K3.1 mark. It is not a picture the model made for this listing.