Kimi K3.1 Pro
Moonshot AI·
LLMsopen-weightcloud
Overview
Post-trained version of Kimi K3 (2.8T MoE, ~50B active) with substantially larger post-training compute, run as large-scale joint RL across hundreds of downstream domains. Launched on Kimi.ai, Kimi Work, Kimi Code and the Kimi API on 2026-10-23, alongside a Highspeed variant (2-3x faster, same capabilities); full weights announced for release by 2026-11-04.
Capabilities and innovations
2.8T total / ~50B active MoE (Kimi K3 base)1M token context windowText and image input3D modeling (Blender, Three.js, game engines)Agentic coding and financial researchOpen weights announced for 2026-11-04Large-scale joint RL across hundreds of downstream domainsHighspeed serving variant at 2-3x throughput
Benchmarks
| Benchmark | Score | Source |
|---|---|---|
| GPQA Diamond | 94.2% | vendor GPQA-Diamond 94.2 at max effort, official launch blog (2026-10-23). |
| Humanity’s Last Exam (no tools) | 51.1% | vendor HLE-Full 51.1 without tools, full 2,500-question set, max effort, recorded for comparability with the no-tools HLE used here (Kimi K3: 43.5). Moonshot also reports 70.0 with tools and HLE-Diamond 39.4, neither comparable to this field. Artificial Analysis measures 52.5 on the 2,158-question text-only subset. |
| Terminal-Bench 4.0 | 63.64% | independent Terminal-Bench 4.0 63.64% (± 1.75), measured by Vals AI, rank 3 of 56 (accessed 2026-10-27). Vals AI is the reference harness for this field so values compare across vendors. Moonshot reports 65.7 from the Claude Code harness; Artificial Analysis measures 67.2 with mini-swe-agent on 66 tasks. 4.0 is its own generation and not comparable with 2.x or 3.0. |
| DeepSWE v1.1 | 75.5% | vendor DeepSWE v1.1 75.5, explicitly labelled v1.1, official launch blog. Moonshot states K3.1 Pro was evaluated with the Claude-Code harness, while DeepSWE scores under a fixed mini-swe-agent harness, so this carries the same harness caveat as the DeepSeek and Muse Spark entries. Vendor-baseline check: the competitor rows in the same table (GPT-6 Astra 74.1, Claude Opus 5.5 74.2, GLM-5.3 66.9, DeepSeek-V4.1-Flash-1017 77.0) match each vendor's own published figure, as the blog's footnote says; they are not entered anywhere from this source. Kimi K3 has no value in the table to compare against. Datacurve's own DeepSWE board was not checked. |
| MMMU-Pro | 84.1% | vendor MMMU-Pro 84.1 (official protocol, images prepended), official launch blog; 90.0 with python is a tool-use setting and not entered. Artificial Analysis measures 83.9, within 0.2 points. |
- SWE-bench Pro: not reported by the vendor — Moonshot reports DeepSWE v1.1, SWE Marathon v1.1, Program Bench and its internal Kimi Code Bench instead of SWE-bench Pro.
- MMLU-Pro: independent evaluation pending — Moonshot publishes no MMLU-Pro value. Vals AI measured Kimi K3 (87.97) but lists no MMLU-Pro run for K3.1 Pro yet (checked 2026-10-27).
- Terminal-Bench 2.x: not reported by the vendor — Moonshot reports Terminal-Bench 4.0 only; no 2.x figure from the vendor or Vals AI.
API pricing
$2 per million input tokens, $10 per million output tokens (USD, cheapest OpenRouter endpoint, checked 2026-10-27)
Architecture and hardware
- Parameters
- 2800B total, 50B active per token (MoE)
Links
More from Moonshot AI
Data curated by AI Model Timeline. See the methodology for admission criteria, benchmark eras and source priority.