kimi-k3
Meant to excel at
-
Positioned for long coding sessions, multi-tool agents and knowledge work; SWE Marathon 42.0 ranks first among compared models.
Vendor claim, relayed smhn.info -
Top vendor-reported scores on browsing and automation: BrowseComp 91.2, Automation Bench, and SpreadsheetBench 2.
Vendor claim, relayed smhn.info -
Debuted first on Arena's Frontend Code leaderboard in blind developer votes, ahead of Claude Fable 5.
Independent evaluator techsy.io -
2.8T MoE with 1M context and image and video input; good for everyday chat, writing and web-app generation.
Independent evaluator mayhemcode.com
Watch out for
-
Trails closed frontier models on some coding evals (DeepSWE 67.5 vs 73.0; FrontierSWE 81.2 vs 86.6).
Vendor claim, relayed smhn.info -
Pricey for an open model ($3/$15 per 1M), always-on reasoning billed as output, and only about 62 tokens/s.
Independent evaluator techsy.io -
Failed about 36% of trap tasks designed around open-weight weaknesses versus 8% for Opus 4.8.
Practitioner report mindstudio.ai -
Weaker than Fable 5 or Opus 4.8 on chart and screenshot reading and multi-file code repair.
Independent evaluator mayhemcode.com -
Current AA index is 44 on the October index version (the earlier 57 figure is from an older version), now below Qwen3.8 Max and GLM-5.3 at 45.
Independent evaluator artificialanalysis.ai -
Artificial Analysis marks Kimi K3 as deprecated and points to Kimi K3.1 Pro, which scores 55 on the same index version.
Independent evaluator artificialanalysis.ai
Benchmarks
Coding percentile 0.45
| Benchmark | Score | Percentile | Source | Date |
|---|---|---|---|---|
| code arena webdev | 1655.2 | 0.43 | Independent evaluator llmrun.dev | undated |
| lmarena code arena webdev | 1655.2 | 0.48 | Independent evaluator lmarena.ai | undated |
Agentic coding percentile 0.76
| Benchmark | Score | Percentile | Source | Date |
|---|---|---|---|---|
| terminal-bench 2.1 | 88.3 | 0.73 | Vendor claim, relayed morphllm.com | 2026-07-27 |
| terminal-bench 2.1 (aggregator, anotherwrapper.com) | 88.3 | 0.79 | Independent evaluator anotherwrapper.com | 2026-10-19 |
Overall percentile 0.26
| Benchmark | Score | Percentile | Source | Date |
|---|---|---|---|---|
| artificial analysis intelligence index | 44 | 0.26 | Independent evaluator artificialanalysis.ai | 2026-10-19 |
Reasoning percentile 0.40
| Benchmark | Score | Percentile | Source | Date |
|---|---|---|---|---|
| gpqa diamond | 93.5 | 0.50 | Vendor claim, relayed morphllm.com | 2026-07-27 |
| humanity's last exam | 43.5 | 0.31 | Vendor claim, relayed morphllm.com | 2026-07-27 |
Percentile is the share of tracked models this model beats on that benchmark, averaged across the group. Only benchmarks at least four tracked models share are scored.
Practitioner reports
- +
Kimi K3 produced the most compact code (about 360 lines per task) in a one-shot HTML animation test versus Qwen3.8 Max at about 540.
Compared with qwen3.8-max | 2026-10-04
Practitioner report empiriolabs.ai - +/-
In a 3D-world generation test via an open harness, K3 self-corrected against a reference image and planned seven sub-agents at a fraction of the cost, but Opus 5 gave the more polished physics and fidelity.
Task: agentic-long-running | Compared with claude-opus-5 | 2026-07-30 older than 90 days
Practitioner report mindstudio.ai - +/-
For web design, K3 produced a working landing page in under 5 minutes and could fix CSS from a screenshot with native vision, but had small spacing/animation errors; Claude's first iteration was closer to production and its auth architecture cleaner.
Task: feature | Compared with claude-opus-4.8 | 2026-07-29 older than 90 days
Practitioner report wesleychong.com - +
Single-prompt tests gave a clear blog intro, an idiomatic React landing page and a faithful summary of a long research paper with no obvious fabrication.
2026-07-28 older than 90 days
Practitioner report chatgoat.ai - +/-
MindStudio's harness on real GitHub issues: K3 stayed within about 0.2 points of Opus 4.8 on simple tasks and above 60/70 on complex builds, but failed 36% of deliberately trapped tasks vs about 8% for Opus; suggests Opus/GPT for planning and K3 for implementation.
Task: feature | Compared with claude-opus-4.8 | 2026-07-25 older than 90 days
Practitioner report mindstudio.ai - +/-
Nearly matched Opus 4.8 on simple GitHub-issue PRs at about half the price, but failed 36% of trap tasks versus 8%; suggests Opus for planning and K3 for implementation.
Compared with claude-opus-4.8 | 2026-07-25 older than 90 days
Practitioner report mindstudio.ai - +
Finished a dataset-build, Gemma fine-tune and local web deploy workflow end to end and fixed its own errors, but did better in Kimi CLI than other harnesses (sponsored review).
Compared with claude-opus-4.8 | 2026-07-25 older than 90 days
Practitioner report volanea.com - +/-
Good for chat, writing and frontend or web app generation, but Fable 5 or Opus 4.8 are better for chart and screenshot reading, high-stakes agents and multi-file repair.
Compared with claude-fable-5 | 2026-07-18 older than 90 days
Practitioner report mayhemcode.com - +/-
In a Rails chat-app build K3 used the correct library API, wrote production-style Docker and failure tests, but skipped a system prompt, rate limits and message-size caps; hardening and test depth trail Opus 4.8. Good for interactive solo coding.
Task: feature | Compared with claude-opus-4.8 | 2026-07-17 older than 90 days
Practitioner report akitaonrails.com - +/-
Willison's pelican-on-a-bicycle SVG test produced a good result but burned 13K reasoning tokens, about 25 cents, for one trivial prompt.
2026-07-16 older than 90 days
Practitioner report techsy.io