Moonshot Stripped the Hacking Data Out of Its Best Model, Kimi K3.1 Pro Matches the Frontier at Half the Price, With Weights Due November 4
The strongest open model yet arrived with a piece deliberately missing
Moonshot AI released Kimi K3.1 Pro on Saturday, October 24, and the headline numbers are the kind that used to belong only to closed labs. It is a 2.8-trillion-parameter mixture-of-experts model with 104 billion parameters active at a time, a 1-million-token context window, and multimodal input. Independent tester Artificial Analysis scored it 55 on its Intelligence Index, ahead of Claude Fable 5.1, GPT-6 Astra and Gemini 4 Argon, each at 53, and behind only Claude Opus 5.5 at 58. On Moonshot’s own reporting, it scores 65.7% on Terminal-Bench 4.0, a hard test of coding agents working in a real terminal, against 66.4% for Opus 5.5 and 58.2% for GPT-6 Astra. Nathan Lambert, who tracks open models closely, called the scores “clearly the strongest current open weight model, if the weights are released as they claim.”
The price is the other half of the pitch: $2 per million input tokens and $10 per million output tokens through the API, the same as OpenAI’s GPT-6.1 Sol and half the price of Claude Opus 5.5. A Highspeed variant runs two to three times faster at twice the price, except for cached input tokens, which cost the same. The model is live today in the Kimi apps, in the Kimi Work desktop app from version 3.1.0, in Kimi Code, and through the API as kimi-k3.1-pro and kimi-k3.1-pro-highspeed, with zero data retention available to API customers. Enterprise access is off by default and has to be switched on by an admin.
And then there is the date that turns an API launch into an industry event: Moonshot has promised to publish the full model weights by November 4. Anyone will be able to download, run and modify the most capable model ever released that way. Which is exactly why the second half of Moonshot’s announcement matters more than the benchmarks.
What Moonshot removed, and the number that proves it cost something
Buried in the technical post is an unusual disclosure. As Moonshot scaled up post-training, the model’s vulnerability-research scores kept climbing. When engineers traced why, they found cyber-related content mixed into the general mid-training data and cyber tasks sitting inside the reinforcement-learning environments. For the open-weight release, they removed both, the cyber data from the training corpus and the cyber tasks from the RL environments, while keeping the general coding tasks that teach a model to write, debug and repair software.
Moonshot published the cost of that decision on its own internal test, xdev-bench-hard. Most of its tasks recreate recently disclosed bugs in real software, the Linux kernel, Node.js, Exim, with protections switched on. The agent gets the vulnerable source code, a short hint and the binary, and passes only if it writes a working exploit that achieves arbitrary code execution in a separate sandbox and extracts a secret string. A checkpoint trained without the filtering scored 90.5%. The model Moonshot actually released scores 81.0%. That 9.5-point gap is, in effect, a receipt: the removal took real capability away, and the company can show by how much, by its own measure.
The stated principle is worth quoting exactly: “We aim not to release open weights that would extend the absolute frontier of cyber capability.” Moonshot adds two qualifications. First, even after the cut, K3.1 Pro’s cyber capability remains, in its words, behind that of frontier proprietary models that have been publicly available for at least three months, the argument being that open weights should not hand attackers anything they cannot already rent from a closed API. Second, the company says it intends to raise the cyber capability in its open weights over time, deliberately, so that more teams can use the models defensively, auditing the code they maintain, including old legacy systems that get little attention, on infrastructure they control.
Why this week, of all weeks, made the argument for it
Read the timing against the last ten days of this site’s coverage and the logic is hard to miss. AI found two critical PHP flaws that had hidden for two decades, and attackers were probing within days. A single hacker used AI agents to break into more than 100 company websites for about $50 an intrusion. China, Moonshot’s home market, proposed export controls last week that would ban exports of AI-driven malware-generation technology and restrict AI-assisted vulnerability-discovery tools, with the draft still open for comment. Reuters reported the Kimi curbs on Saturday under a headline about fears that increasingly autonomous AI could turbocharge hacking, and those fears are no longer abstract.
The Western labs have been making the mirror-image case. Anthropic argued in a report in late September that an open model from another Chinese lab, Z.ai’s GLM-5.3, posed a major cyberattack risk precisely because anyone can download and modify it. OpenAI’s CFO Sarah Friar told Bloomberg Television on October 14: “We think you want to be on more Western models on the cyber front.” Sam Altman has said OpenAI’s best model cannot responsibly be open-sourced. Moonshot is trying to occupy the ground between those positions: open weights are coming, at frontier level, but the releaser accepts a duty to ship them below the cyber frontier, and publishes a number to prove it tried.
The honest caveats
Three of them, and they matter.
- The benchmarks are mostly the vendor’s. Moonshot’s coding scores come with footnotes: K3.1 Pro ran inside Claude Code for most coding tests, while many rival scores were taken from public leaderboards or the rivals’ own reports. Artificial Analysis’s independent 55 is the strongest external signal so far, and even it will be re-tested once the weights are public and anyone can run the model without Moonshot’s serving stack in the loop. Treat November 4, not today, as the verification date.
- Training-time removal is not a lock. Moonshot’s controls apply to what went into training. Its post does not address the obvious follow-up: once the weights are public, someone can fine-tune the model on cyber data and try to put the filtered capability back. General coding skill already carries over, that is why the filtered model still scores 81.0%, and fine-tuning a 104-billion-active-parameter model is expensive but well within reach of the organisations defenders worry about most. The curb raises the cost and the effort; it does not settle the question.
- The finance pitch needs the same discount. Moonshot is selling K3.1 Pro to finance teams on an internal stress test where its claim-level hallucination rate fell from 23.4% on K3 to 8.3%. The company itself adds that “the reported rates do not represent overall error rates in daily use.” Credit for printing the caveat; do not build a compliance case on the number.
What to do with this, this week
- If you run agents on API models, price-test it now. The API is live, the price is half of Opus 5.5, and the independent score is within three points. Run your own tasks, not a public benchmark, through kimi-k3.1-pro alongside your current model and compare quality per dollar on work you actually ship.
- If you plan to self-host, plan for November 4 honestly. A 2.8-trillion-parameter model, even with 104 billion active, is not a laptop model. Budget for serious GPU infrastructure or a hosted open-weight provider, and do not commit until the weights, the licence and independent benchmarks are all public.
- If you are on a security team, add it to the threat model either way. A frontier-class coder scoring 81% on an internal exploit suite will be downloadable in eleven days. That is a reason to accelerate patching and code audit, including using models like this one defensively, which is the use Moonshot says it wants to enable over time.
- Watch what “responsible open release” becomes. If Moonshot’s approach, publish the capability you removed, publish the score it cost, state the frontier you refuse to extend, becomes the norm other open labs are judged against, Saturday will matter long after the benchmark lead changes hands.
The bottom line
For two years the open-weights debate has been a shouting match between “release everything” and “release nothing,” with no shared evidence in the middle. Kimi K3.1 Pro does not resolve that debate, but it changes its terms: a Chinese lab shipped the strongest soon-to-be-open model yet, at half the leading closed price, and voluntarily weakened the one capability the whole industry spent October worrying about, then published the before-and-after numbers. Whether 81.0% is low enough, and whether fine-tuning undoes the work, are questions the next month will answer in public. That, in the end, is the point of open weights: everyone gets to check.