Models › Text › GPT-5.6 Sol
Moonshot AI • Launched July 2026 as the largest open-weight model
Kimi K3 is Moonshot AI’s largest open-weight reasoning model yet
Starting provider API rate; context tiers apply
See what K3 does best, then review its limits, benchmarks, and pricing below
Kimi K3 is Moonshot AI’s largest open-weight reasoning model yet
Every K3 result below comes from a named source and is labeled by protocol, so each score reflects how it was measured
| Benchmark | Score | Evaluation details | Source |
|---|---|---|---|
GPQA Diamond The highest-quality subset of Graduate-Level Google-Proof Q&A, designed to test expert-level scientific reasoning in biology, physics, and chemistry. |
93.5% Accuracy |
Version: GPQA Diamond 198 · Split: 198 Diamond questions · Harness: Artificial Analysis independent evaluation · Reasoning effort: max | Methodology Artificial Analysis Independently reproduced |
Humanity's Last Exam A 2,500-question expert-level benchmark spanning dozens of academic fields. Tool-assisted and no-tools results are separate protocols. |
43.5% Accuracy (no tools) |
Version: Humanity's Last Exam, full set · Protocol: No external tools · Harness: Artificial Analysis independent evaluation · Reasoning effort: max | Methodology Artificial Analysis Independently reproduced |
SWE-bench Verified A 500-task benchmark of real GitHub issues, graded by whether the generated patch passes hidden unit tests. |
76.8% Pass@1 |
Version: SWE-bench Verified · Split: 500 verified tasks · Harness: Kimi Code (Moonshot's own agent harness) · Reasoning effort: max | Methodology Moonshot AI Provider-reported |
Terminal-Bench 2.1 A benchmark of realistic terminal-based tasks that measures whether an agent can complete real command-line workflows end to end. |
88.3% Task success rate |
Version: Terminal-Bench 2.1 · Harness: Kimi Code (Moonshot's own agent harness) · Reasoning effort: max | Methodology Moonshot AI Provider-reported |
DeepSWE 1.1 A long-horizon software-engineering benchmark with 113 original tasks graded by hand-written tests. |
67.3% Pass@1 |
Version: DeepSWE 1.1 · Split: 113 tasks across 91 repositories · Harness: mini-swe-agent · Reasoning effort: max | Methodology DataCurve Benchmark owner |
Arena rankings capture how people vote between anonymous responses. K3’s human preference standing is reported per leaderboard, since results can vary widely by task
K3 ranked first on LMArena’s Frontend Code leaderboard for building working front-end interfaces, scoring 1,679 Elo, a 17-place jump from its predecessor
Kimi K3 runs a 1,048,576-token context window at one flat rate, with image input, always-on thinking, and open weights, priced below
| Model | |
| Provider | Moonshot AI |
| Release date | Jul 16, 2026 (API) • Jul 27, 2026 (open weights) |
| Catalog added | Sep 22, 2026 |
| Context window | 1,048,576 tokens |
| Maximum output | 131,072 tokens |
| Knowledge cutoff | Not available in reviewed sources |
| License | Kimi K3 License (open-weight, modified MIT) |
| Capabilities and API | |
| Parameters | 2.8T total • 104B active (MoE, 896 experts) |
| Input types | Image, Text |
| Output types | Text |
| Reasoning | Always-on thinking mode (max effort by default) |
| Provider endpoint tool parameters | Accepts tool parameters |
| Base provider API input | $3.00 / 1M tokens |
| Base provider cached input | $0.30 / 1M tokens |
| Base provider API output | $15.00 / 1M tokens |
Provider token rates are context-dependent. The base rates apply below the listed prompt thresholds; the matching tier applies at or above each threshold. These are reference provider API rates, not Pyxa AI checkout charges — access in Pyxa AI follows the allowances of your Pyxa AI plan.
Each K3 benchmark and specification links to its original source, so you can check where every figure comes from
Discover where Kimi K3 performs best, from long-horizon coding and research to open-weight self-hosting
Put Kimi K3 to work on your coding, research, and long-context tasks inside Pyxa AI