"The only way to discover the limits of the possible is to take a small step beyond the impossible." — Moonshot AI
On October 24, Moonshot AI (Kimi) released Kimi K3.1 Pro, its new open-source flagship: 3 trillion parameters, a 1-million-token context window — still the largest open model on Earth. Its overall intelligence now approaches leading closed-source models, at one-fifth the price of GPT 6 Astra. The open-versus-closed contest has entered a new round.
1. Release and Positioning
On the Artificial Analysis intelligence index, the K3 family ranks 4th overall — first among open models worldwide. Versus K3, K3.1 Pro pours far more compute into reinforcement learning across hundreds of domains and wields professional tools far more capably — growing more "hands" beyond its intelligent "brain". On coding, agentic and knowledge benchmarks it matches or approaches the strongest closed models.
Key Facts at a Glance
• Released Oct 24: 3T parameters, 1M-token context — the world's largest open model
• Priced at one-fifth of GPT 6 Astra; full weights open by Nov 4
• Ranks 4th on the Artificial Analysis index, first among open models
• Long-horizon coding: a simulated human walked for 11.2s, with 36 moment-arm errors auto-fixed
• Global best on both Office QA Pro v2 and SpreadsheetBench v2
2. Three Leaps: Coding, 3D and Finance
- ① Coding: In an internal test, the model built a simulated human walking system — iterating from collapsing within a few steps to 11.2 seconds of continuous walking, fixing 36 errors in shoulder muscle moment-arm calculations and retraining the controller.
- ② 3D creation: From a text prompt, a sketch or a reference image it infers shape, proportion and spatial relations, driving professional modeling software to produce interactive scenes — strong at mechanical parts, architectural models and custom furniture, previewable and editable online, downloadable for 3D printing.
- ③ Finance: End-to-end delivery from retrieval and cross-document verification to financial modeling, target screening and earnings analysis; it takes the global best on both Office QA Pro v2 and SpreadsheetBench v2, ahead of Opus 5.5, Fable 5.1 and GPT 6 Astra.
3. The Open-Source Timeline
- ① Backstory: Kimi K3, released July 16 at 2.8 trillion parameters, was then the world's largest open model — within 48 hours user requests pushed its cluster near its limits; it is now fully open for download and deployment.
- ② Cadence: Full weights of K3.1 Pro are due by November 4; per reports, a Highspeed variant runs 2–3× faster with the same million-token context.
- ③ Roadmap: Founder Yang Zhilin said at the AGI-Next summit earlier this year that K4, K5 and up to a K100 series will roll out over the coming decade or two.
From K3's instant capacity crunch to K3.1 Pro closing on closed rivals at one-fifth the price, open-source is pushing the contest into a new phase of "capability parity, price descent". As open weights and million-token contexts become table stakes, the decisive battleground shifts to deployment — and the private-deployment ledger is one every enterprise eyeing large models should run carefully.
(Compiled from The Paper, HK01 and other public reports.)