Kimi K3.1 Pro
B · Useful, limited Moonshot AI's release post for a large open-weight model, with detailed benchmark footnotes and an unusual cyber-capability ablation; scores are vendor-run or vendor-selected and the weights were not yet released.
Summary based on the full text·Original: about 15 minutes to read
In short
A Moonshot AI technical blog post introducing Kimi K3.1 Pro, a post-trained version of Kimi K3 that the company calls its most capable 3T-class open-weight model, with a 1-million-token context window. Moonshot reports scores close to the strongest proprietary models on coding, 3D/CAD and finance benchmarks. The most distinctive section is on cybersecurity: Moonshot removed cyber-related data from mid-training and RL and shows the capability this cost, as the basis for its open-weight release decision. The page is undated; weights are promised by 4 November 2026.
Key numbers
- 65.7% vs 66.4%Terminal-Bench 4.0, K3.1 Pro vs Claude Opus 5.5 (GPT-6 Astra 58.2%), per Moonshot
- 90.5% vs 81.0%xdev-bench-hard exploit tasks solved at pass@3, checkpoint without cyber-data removal vs released model
- 23.4% to 8.3%claim-level hallucination rate on Moonshot's internal stress test, K3 to K3.1 Pro
- 1-million-tokencontext window
- by Nov 4date by which full model weights are promised
As reported by the source.
English·Web page·opens kimi.ai
What it reports
- Terminal-Bench 4.0: 65.7%, against GPT-6 Astra 58.2% and Claude Opus 5.5 66.4%. Internal Kimi Code Bench 3.0 Long-horizon: 61.2% vs K3's 55.95%.
- BenchCAD (CAD code from multi-view renders, with tools): 92.7%, against Opus 5.5 at 96.2% and GPT-6 Astra at 95.9%.
- OfficeQA Pro v2: 64.4, which Moonshot calls state of the art. Internal finance benchmarks: +7.2 points on financial modelling, +11.6 on financial research, about 4.2x K3 on a business simulation.
- Internal hallucination stress test: claim-level rate fell from 23.4% to 8.3%; Moonshot notes this does not represent everyday error rates.
- Cyber: without data removal, a checkpoint solved 90.5% of xdev-bench-hard exploit tasks at pass@3; the released model solves 81.0%. Moonshot says this stays behind proprietary frontier models public for at least three months, and states it aims not to extend the open-weight cyber frontier while raising it gradually.
- A Highspeed variant runs 2-3x faster. Enterprise access is off by default at launch.
How they got it
Vendor evaluations, mostly run through the Claude Code harness at maximum reasoning effort; competitor scores are taken from their reports or leaderboards, sometimes as the maximum reported value. Footnotes disclose substitutions (for example H20 instead of H100 GPUs, zero scores for runs that fetched answers online). xdev-bench-hard and several finance and hallucination tests are internal and unpublished.
Limits
- Mixed sources for competitor scores and harness differences make head-to-head comparisons loose.
- Key safety and accuracy claims rest on internal benchmarks outsiders cannot run.
- Weights were promised, not released, at the time of the post, so the open-weight claims could not yet be checked.
- The page carries no date.
Why it matters for companies
A near-frontier open-weight model changes the build-or-buy math for firms that need to run models on their own infrastructure for data or cost reasons. Run your own task evaluations before switching, since vendor benchmarks are selected. The cyber section is worth reading for security teams: it shows a developer deliberately trading capability for lower misuse risk and measuring the trade, which is the kind of disclosure to ask of any model supplier.
Summary based on the full text. Read the original for the full argument and data.Read the original on kimi.ai
- Tags
- Open weights