Moonshot's new flagship goes live today with a million-token context and a promise to publish full weights by November 4. The part worth reading is what the company took out of the model before release.
Kimi K3.1 Pro is the first model in Moonshot's K3.1 series
Moonshot AI released Kimi K3.1 Pro on Friday, October 23. It's the first model in a new K3.1 line, and the company calls it its most capable 3T-class open-weight model so far.
The "3T-class" label is Moonshot's own. MarkTechPost puts the size at 2.8 trillion total parameters with 104 billion active per token, which fits a mixture-of-experts design where only part of the network runs on any given token. Treat the exact figure as reporting, not a spec sheet.
Under the hood, K3.1 Pro is a post-trained version of Kimi K3. Same base, much more post-training compute. Moonshot says it rebuilt the reinforcement learning stack, long-context inference, sandbox infrastructure and rollout scheduling so that extra compute produced steady gains instead of noise.
A second model ships alongside it. Kimi K3.1 Pro Highspeed runs two to three times faster for interactive work.
Both are live today on Kimi.com, the Kimi mobile apps, Kimi Work, Kimi Code and the Kimi API. Enterprise admins can turn K3.1 Pro on for their organization, off by default.
The weights are the exception. Moonshot says full weights will be released by November 4, 2026. Until then, "open" describes the plan, not the download.
What a 1M-token context window actually buys you
Kimi K3.1 Pro takes 1,048,576 tokens of context. A million tokens, in marketing terms, and it's the number Moonshot wants you to remember.
For a developer, it means a mid-sized repository can go into one prompt without a retrieval pipeline bolted on top. For an analyst, it means a full annual report plus the spreadsheets behind it. No chunking. No re-uploading the same PDF four times.
Does that matter in daily use? Sometimes. Long context is easy to advertise and harder to use well, because retrieval quality tends to sag past the middle of a very long prompt. Moonshot claims gains there but hasn't published a needle-in-a-haystack breakdown by position.
Kimi K3.1 Pro benchmarks, with the company asterisk attached
Every score below comes from Moonshot's own blog. None has been independently reproduced yet.
On Terminal-Bench 4.0, which tests coding agents on realistic command-line tasks with strict per-task checks, K3.1 Pro scores 65.7%. GPT-6 Astra sits at 58.2% and Claude Opus 5.5 at 66.4% in the same table.
That puts K3.1 Pro ahead of one closed frontier model and a hair behind another. On an open-weight release, that's the headline.
On Kimi Code Bench 3.0 Long-horizon, an internal test of multi-step coding, it scores 61.2% against K3's 55.95%.
BenchCAD, which asks a model to turn multi-view renders into CAD code, gives 92.7% with tools. Claude Opus 5.5 reaches 96.2% and GPT-6 Astra 95.9%, so this is a case of closing a gap rather than setting a record.
OfficeQA Pro v2 lands at 64.4, which Moonshot describes as state of the art for fiscal document analysis.
The pattern is consistent. Strong, close to the leaders, and self-reported. Ask again in a month, when outside testers have run the weights.
Kimi K3.1 Pro's hallucination numbers need the fine print
Moonshot's most quotable figure is a hallucination rate falling from 23.4% on K3 to 8.3% on K3.1 Pro. Claim-level, on an internal test called Kimi Hallu Bench Hard.
Here's the caveat, and Moonshot states it plainly. The benchmark is a stress test. Its examples were chosen to trip models up on purpose. The reported rates aren't daily error rates, and the blog says so.
So the useful reading is relative, not absolute. The same adversarial test that caught K3 on 23.4% of claims catches K3.1 Pro on 8.3%. That's a big improvement on a test designed to be hard.
Moonshot made K3.1 Pro worse at hacking, on purpose
This is the unusual part of the release, and it deserves more attention than the benchmark table.
Moonshot says it found cyber-related content inside its general mid-training corpus and cyber tasks inside its RL environment mix. Left alone, those pushed vulnerability research scores up as post-training scaled. So the company removed that data from both stages for the open-weight release, keeping general coding tasks intact.
The measurement: on xdev-bench-hard, an in-house set of exploit-writing tasks that require arbitrary code execution in a separate sandbox, an untreated checkpoint solves 90.5% at pass@3. Final K3.1 Pro solves 81.0%.
That's a 9.5-point cut, produced deliberately.
The cut is disclosed in the blog post, which is more than most labs do.
Moonshot also says K3.1 Pro's offensive cyber capability stays below what its pipeline can produce, and behind frontier proprietary models that have been public for at least three months. Moonshot argues that an open model running on hardware its users control lets far more teams audit legacy code that rarely gets a dedicated security review.
You can accept the reasoning and still note the gap. These are Moonshot's own tests, not an outside audit.
Kimi K3.1 Pro API pricing and the Highspeed option
Moonshot's blog doesn't publish per-token rates. Third-party listings do: about $2 per million input tokens, $10 per million output, and $0.20 for cached input. Highspeed is listed at roughly double, $4 and $20.
Treat both as a market signal rather than a contract, since aggregator pages drift and API pricing moves with little notice.
For agent work, the cached-input rate matters more than the headline. Agents resend the same system prompt and tool definitions on every step of a long run, which means cache pricing often decides whether the work is affordable at all.
At $2 input, K3.1 Pro sits below Claude Opus 5.5 and above the cheapest Chinese models.
Cheap, for its class.
That's the slot Moonshot is aiming for: frontier-adjacent quality at a price that won't end a pilot.
What to check before Moonshot's open-weight release on November 4
Three things will decide whether this lands as well as the benchmarks suggest.
The license, first. Moonshot hasn't published terms for K3.1 Pro yet, and an open-weight model with tight commercial strings is a different product from one without them.
Second, whether the cyber filtering survives contact with fine-tuning. A base model that scores 81.0% on exploit tasks can be pushed back up by anyone with compute and intent. Moonshot is betting that gradual diffusion helps defenders more than it helps attackers. That bet is testable.
Third, independent evaluation. Terminal-Bench numbers from a company blog are a claim. Terminal-Bench numbers from a third party with the weights in hand are evidence.
If you want to try K3.1 Pro now, it's in the model picker on Kimi.com, Kimi Work and Kimi Code, and behind the kimi-k3.1-pro model ID on the API. The weights are the thing to set a reminder for.
