Moonshot releases Kimi K3.1 Pro
2.8T MoE, 1M context, 4th on Artificial Analysis behind only Anthropic and Meta, weights promised by Nov 4, cyber data deliberately removed
Confirmed
Importance: major (4 of 5)The takeaway
On Oct 23, 2026 Moonshot AI launched Kimi K3.1 Pro (2.8T total / 104B active MoE, 1M context, multimodal) and a 2–3x faster Highspeed variant on Kimi.ai, Kimi Code and its API at $2/$10 per 1M tokens, with full weights promised by Nov 4.
Status
- Claim
Confirmed
- Our reporting
- High confidence
- Importance
- Major (4 of 5)
- Last verified
- 25 October 2026
Your AI and this story
- GPT-6 Astra176 days after its cutoff
- Claude Opus 5.5115 days after its cutoff
- Gemini 4 Argon206 days after its cutoff
- Grok 4.7145 days after its cutoff
None of these four assistants can know about it. The closest, Claude Opus 5.5, stops 115 days before it.
Key facts
- Architecture: 2.8T total / 104B active parameters (MoE), 1M-token context, multimodal input (Moonshot on X, 2.58M views)
- API ids: kimi-k3.1-pro and kimi-k3.1-pro-highspeed; OpenRouter moonshotai/kimi-k3.1-pro at $2 input / $10 output / $0.20 cached per 1M tokens (Willison: half Opus 5.5’s price; Highspeed about twice the price)
- Artificial Analysis: Intelligence Index 55 (K3: 44), 4th behind Opus 5.5 (58), Sonnet 5.5 (56) and Muse Wonder Full (56); GDPval-AA 1,846 Elo (3rd, +312 vs K3); Terminal-Bench 4.0 67%, the highest AA had measured (K3: 13%); AutomationBench-AA 58% → 73%; HLE 47% → 52%
- Moonshot’s own numbers: Terminal-Bench 4.0 65.7% (Opus 5.5 66.4%, GPT-6 Astra 58.2%); BenchCAD with tools 92.7% Voxel IoU; OfficeQA Pro v2 64.4; claim-level hallucination rate on its hard benchmark 23.4% (K3) → 8.3%
- Cyber data removal (blog): cyber content was found embedded in the general mid-training corpus and in the RL mix; ‘we identified and removed cyber-related data from both training stages’. Goal: not to release open weights ‘that would extend the absolute frontier of cyber capability’. On xdev-bench-hard the released model scores 81.0 vs 90.5 for the unfiltered checkpoint (GPT-5.6 Sol 100)
- Moonshot says K3.1 Pro’s cyber capability trails frontier proprietary models that have been public for at least three months, and plans to ‘progressively raise’ the cyber capability available in open weights
- Marketed for full-project game coding in Unreal, Unity and Godot, parametric 3D/CAD in Blender, and finance modelling
- Enterprise admins must switch it on (off by default); Zero Data Retention offered on the API
Show 2 more
- Hacker News: 927 points, the day’s top AI story
- Influencer allegations (VentureBeat, Oct 24): X user @elshayib_ posted screenshots of what he called a Kimi promotion brief asking creators to submit drafts for approval, with the line ‘Please DO NOT put on the paid partnership label’ and talking points casting Kimi as an open Claude Opus 5.5 at a fraction of the price; the document is not authenticated, its issuer and reach are unknown, and no Moonshot response was found as of Oct 25. VentureBeat notes this does not affect the independent Artificial Analysis results (Intelligence Index 44 → 55 from K3)
What happened
Moonshot released K3.1 Pro three months after Kimi K3. It is served first through its own apps and API, with the weights to follow by Nov 4 (K3’s weights followed 11 days after launch). Artificial Analysis ranked it just behind Anthropic’s two top models and Meta’s Muse Wonder Full. It leads every OpenAI and Google model on that index and posts the best Terminal-Bench 4.0 score AA has measured.
The launch post also explains a safety choice. Moonshot found that vulnerability-research skill rose steadily as post-training scaled, because cyber material was mixed into general data and RL tasks. It filtered this data out of both stages, and the released model gives up about 10 points on its internal exploit benchmark.
Why it matters
Once the weights ship, a near-frontier model will be free to download, three points behind the best closed model on a broad index. A Chinese lab is also publicly holding back offensive cyber capability in an open release. Western labs have asked for this, and it comes the same day OpenAI took all its frontier training offline over cyber-capable models. Whether the filtering survives fine-tuning by others is untested.
Sources
9 sources from 7 sites. Numbers match the chips in the text.
9 sources: 3 primary, 2 press, 4 reactions
Primary
- Kimi blog: Kimi K3.1 Prokimi.ai, official
- Kimi.ai on X: Introducing Kimi K3.1 Pro (2.58M views)x.com, official
- OpenRouter: moonshotai/kimi-k3.1-proopenrouter.ai, docs
Press
- VentureBeat: Kimi K3.1 Pro flirts with open-weights leadership amid paid influencer scandal allegations (Oct 24)venturebeat.com, press
- The Decoder: Chinese AI model Kimi K3.1 Pro is closing in on Anthropic’s Claude Opus 5.5the-decoder.com, press
Reactions
- Artificial Analysis on X: Kimi K3.1 Pro launches 4th on the Intelligence Indexx.com, discussion
- Simon Willison: Introducing Kimi K3.1 Prosimonwillison.net, discussion
- Andrew Curran on X: the cyber-data passage of the launch postx.com, discussion
- Hacker News discussion (927 points)news.ycombinator.com, discussion
Changes
- Added the undisclosed-influencer-promotion allegations (VentureBeat)
- Filed (lab X posts, Hacker News; launch blog read directly). Resolves the K3.1 calendar item