Hands-On With Kimi K3.1Pro: Open 2.8T-Param Flagship Halves Price, Matches GPT on 3D
数字生命卡兹克 · wechat · 2026-10-24
A detailed overnight review of Moonshot's new Kimi K3.1Pro: a 2.8T-parameter, 1M-context open flagship (weights due by Nov 4) with API pricing cut 50% vs K3 — roughly one-fifth of GPT-6Astra's rate.
- Benchmarks: 65.7% on Terminal-Bench 4.0 (beats GPT-6Astra's 58.2%, just under Opus 5.5's 66.4%); hallucination rate down from 23.4% to 8.3%; SOTA on OfficeQA Pro v2.
- Real-world tests: 3D output nearly matches GPT-6Astra — jelly physics, a high-precision mechanical keyboard build, and a walkable Three.js home generated from a floor plan plus a decor reference image, even reconciling a handheld video to add furniture missing from the plan.
- Motion/video: one-prompt motion-design openers, narrated promos, and a near-30-minute explainer video; still behind Opus 5.5 but a big leap.
- Includes full prompts and a multi-agent iteration workflow for the 3D house task.
Related event: Kimi K3.1 Pro hands-on: strong visuals, mixed general performance(2 posts)→
More from coding & agent
- Alchemy author: a software factory is conventions, not just agent orchestration — samgoodwin89 · 2026-10-24
- Claude Code spotted a malicious extension stealing cookies, saving $6,800 in crypto — Arindam_1729 · 2026-10-24
- Creator hooks ChatGPT to Runway via MCP to drive Seedance 2.5 cinematic clips — aziz4ai · 2026-10-24
- Kimi K3.1 Pro tested: 2-3x faster at $0.55 per task, open models close on frontier 3D — cedric_chee · 2026-10-24
- Dev builds Supply Drop, a Grok-powered disaster recovery assistant with four practical desks — seanmcdonaldxyz · 2026-10-24
- LiteLLM 1.107.0 supply chain compromise: credential theft triggered by just starting Python — yuridiogenes · 2026-10-24