Moonshot AI releases Kimi K3.1 Pro, a 3T-class open-weight model with 1M context; weights due by November 4

Moonshot AI released Kimi K3.1 Pro on October 23: a 3T-class reasoning model with a 1M-token context window, the first model in the company’s new K3.1 series and, by its own description, its most capable 3T-class open-weight model to date. The announcement carries one number worth checking against: 65.7% on Terminal-Bench 4.0, within 0.7 points of Claude Opus 5.5’s 66.4% and well clear of GPT-6 Astra’s 58.2%. The API is live today under the model ID kimi-k3.1-pro, but the weights themselves are only promised — “released by Nov 4, 2026” — and as of this writing there is no official Hugging Face repo, so self-hosters still have to wait.

Key points

• Terminal-Bench 4.0: 65.7% (vendor-run, in the Claude Code harness), 0.7 points behind Opus 5.5’s 66.4%. On Kimi Code Bench 3.0 Long-horizon, K3.1 Pro scores 61.2% against 55.95% for its predecessor K3.
• Pricing is not on the official blog. OpenRouter and Simon Willison both list $2 per million input tokens and $10 per million output; a Highspeed variant costs $4/$20 and runs, per Moonshot, 2–3× faster.
• Weights are a promise, not a download: the blog says full model weights will be released by November 4, 2026, so for now the API is the only way in.

Background

Kimi K3 already pushed the open-weight ceiling once, at the end of September. Two threads from our earlier coverage matter here: Inferact — the startup founded by the original vLLM team — ran K3 on 16 TPU v7 chips 57% faster than an equal GB200 setup, evidence that open-weight models had become the standard benchmark workload for inference vendors; and the OpenAI distillation dispute of October 1 turned “how did Kimi catch up?” into an open question. K3.1 Pro shifts the target: the blog introduces it as the first model of the K3.1 series and Moonshot’s most capable 3T-class open-weight model yet, and the comparison set in Moonshot’s own tables is Opus 5.5 and GPT-6 Astra — the current frontier — rather than K3’s own past scores.

The release also closed a rumor loop within hours. Early on October 23 a stealth model codenamed Kibitz appeared on LMArena, and our digest that day logged it as user-side speculation Moonshot had not confirmed. The official blog went up the same afternoon.

The facts

Numbers below come from the official blog of October 23; the stated evaluation setup is reasoning effort max, temperature 1.0, top-p 0.95.

Benchmark	K3.1 Pro	Comparison
Terminal-Bench 4.0	65.7%	Opus 5.5 66.4%, GPT-6 Astra 58.2%
Kimi Code Bench 3.0 Long-horizon	61.2%	K3: 55.95%
BenchCAD (with tools)	92.7%	Opus 5.5 96.2%, GPT-6 Astra 95.9%
OfficeQA Pro v2	64.4	best on its board, per Moonshot
Kimi Hallu Bench Hard hallucination rate	8.3%	K3: 23.4%

Some of the claims need their own attribution lines:

• Parameter count: the blog says only “3T-class”. The finer breakdown — 2.8T total parameters, 104B active, MoE — comes from Kimi’s X announcement as relayed by OpenRouter’s model page. Two different sources, two different levels of precision; don’t merge them.
• Context and modalities: 1M tokens of context. OpenRouter’s page lists text, image and video inputs, with a maximum completion of 943,718 tokens.
• Availability: live on Kimi.com, Kimi Work, Kimi Code and the Kimi API; the Kimi Work desktop app needs v3.1.0 or later on Windows and Apple-silicon Macs. Enterprise access is off by default at launch.
• Security scrub: Moonshot says it identified and removed cyber-offense-related data in mid-training and RL ahead of the open-weight release, and published the cost of doing so — the uncleaned checkpoint scores 90.5% on xdev-bench-hard pass@3, versus 81.0% for the cleaned K3.1 Pro.
• Zero Data Retention is offered on the API.

What others say

Simon Willison tested the API the same day, using the llm openai endpoint command added in LLM 0.32 to hit api.moonshot.ai/v1 directly. His test is the now-traditional pelican riding a bicycle as SVG, generated at reasoning effort low, medium and max, with fresh K3 comparisons at the same levels. His one-line verdict is the sharpest summary of the release: this is the first time a (soon-to-be) open-weight model has scored this highly against the frontier. He also cites Artificial Analysis’s independent score — 55 for K3.1 Pro, ahead of Claude Fable 5.1, GPT-6 Astra and Gemini 4 Argon at 53, behind Opus 5.5 at 58 (via Willison; we did not verify the original Artificial Analysis page). On price, he supplies the context Moonshot left out: input matches GPT-6.1 Sol and runs about half of Opus 5.5’s rate. OpenRouter’s announcement at 18:31 UTC confirmed both variants were live on its marketplace.

The Hacker News thread (story 50212065) climbed all through publication day: the 20:57Z Wayback snapshot already recorded 498 points, and the UTC October 24, 01:00 snapshot shows 700, with comment counts past 360. The discussion split along predictable lines. Practitioners brought field reports: user dom96 said K3.1 Pro, working on his language-interpreter project, “managed to deduce the meaning of the hack statement in the language better than Opus 5.5 did”. The open-weights promise drew the sharpest argument. One camp called the security scrub theater, arguing that anyone who gets the weights can fine-tune the removed capabilities back in; the pushback was that this beats refusal training, which “can be trivially abliterated out” — anyone who truly wants the offensive capability has to train the 3T weights themselves. zozbot234 added a calibration: this is only a post-trained checkpoint of the K3 pre-trained model, and the scrub simply leaves offensive cyber out of the post-training objectives. And one comment cut through the week’s “pacing the frontier” debate: if such a pact exists, Chinese labs never signed it.

Our take

For developers the practical draw is the combination, not any single number: $2/$10 pricing, 1M context, and a Highspeed variant that makes long-context agent loops cheaper to run. If the weights arrive on schedule, November 4 turns K3.1 Pro into the next must-run workload for inference shops — the Inferact TPU benchmarks will almost certainly get a sequel.

Two caveats before you re-plan around it. Until the weights land, every score describes the API version only. And the gap between Moonshot’s own benchmarks (Kimi Code Bench is a Kimi-named benchmark) and the independent 55 from Artificial Analysis is worth holding in mind; wait for that page to be verified on its own before making a purchasing decision on it. The 23.4% → 8.3% hallucination improvement is likewise a vendor-run number until someone else reproduces it.