Kimi K3 benchmarks
Kimi K3 has 7 published scores in this dataset, from a Moonshot AI release dated 2026-07-27. Its composite index is 90.6, which places it 7th of 22 ranked models and 3rd among 15 open-weight models. Its strongest area is coding (index 97.5) and its weakest is reasoning (index 82.4). Its best placing is on Terminal-Bench, where its 88.3 ranks 3rd of 13 models with a published score.
OpenLLM Buddy runs Kimi-K3-Uncensored-GGUF, an uncensored fine-tune of Kimi K3, on 8× H100 80GB (640 GB VRAM). The scores below are the official base model's; a fine-tune can score slightly differently.
Category profile
Index 0–100: each score relative to the best published score on that test.
Model facts
- Lab
- Moonshot AI
- Type
- OpenLLM Buddy catalog
- Released
- 2026-07-27
- Parameters
- 2.8T total, 104B active (MoE)
- Context
- 1M
- License
- Kimi K3 License (custom)
- Runs on
- 8× H100 80GB
Kimi K3 hardware requirements and cost
Even at 4-bit it needs about 1.9 TB (Multiple nodes (over 1.1 TB)), well beyond a home workstation, so most people run it on rented multi-GPU servers. Full BF16 precision needs about 6.4 TB (Multiple nodes (over 1.1 TB)).
| Precision | Weights | With headroom | Smallest setup |
|---|---|---|---|
| BF16 (full precision) | 5.6 TB | 6.4 TB | Multiple nodes (over 1.1 TB) |
| FP8 / 8-bit | 2.8 TB | 3.2 TB | Multiple nodes (over 1.1 TB) |
| 4-bit (GGUF Q4_K_M) | 1.7 TB | 1.9 TB | Multiple nodes (over 1.1 TB) |
Estimate: parameters × bytes per parameter, plus 15% for the KV cache and runtime at modest context lengths. Long contexts and large batches need more.
On OpenLLM Buddy: Kimi-K3-Uncensored-GGUF runs on 8× H100 80GB as a GGUF quantized build, with an OpenAI-compatible API and no per-token fees. Prices start at $422 USD for 11 hours.
- 11 hours: $422 USD
- 24 hours: $920 USD
- 1 week: $6440 USD
- 1 month: $25759 USD
Lighter or stronger open models
Open models ranked just above and below on the composite index, with their 4-bit memory needs.
- Kimi K3.1 Pro94.4—
- DeepSeek V4 Pro91.01.1 TBcompare
- Tencent Hy4 Preview89.9531 GB
- Muse Spark Open88.0563 GB
- Reflection Beam 1.087.0346 GB
Kimi K3 speed and pricing
Measured figures link to their source; estimates are labelled.
- API output speed
- 66 tok/sArtificial Analysis median · Artificial Analysis
- Time to first token
- 1.20 sOpenRouter model page provider stats
- API price per 1M tokens
- $3.00 in · $15.00 outlab's own endpoint on OpenRouter · OpenRouter endpoints API
- OpenLLM Buddy GPU rental
- from $38.33 per hourflat rate, no per-token fees
Score by test
Rank is among every model with a published score on that test.
Nearby in the composite ranking
Models with the closest composite index.
Head-to-head comparisons
Side-by-side pages built from the tests both models report.
Sources
| Model | Test | Score | Setting | Source |
|---|---|---|---|---|
| Kimi K3 | GPQA Diamond | 93.5 | max reasoning effort | Kimi K3 model card (Hugging Face) |
| Kimi K3 | Humanity's Last Exam | 43.5 | HLE-Full, no tools, max effort | Kimi K3 model card (Hugging Face) |
| Kimi K3 | Humanity's Last Exam (with tools) | 56.0 | HLE-Full, with tools, max effort | Kimi K3 model card (Hugging Face) |
| Kimi K3 | Terminal-Bench | 88.3 | Terminal-Bench 2.1, max effort | Kimi K3 model card (Hugging Face) |
| Kimi K3 | BrowseComp | 91.2 | max effort | Kimi K3 model card (Hugging Face) |
| Kimi K3 | MMMU-Pro | 81.6 | MMMU-Pro, no tools | Kimi K3 model card (Hugging Face) |
| Kimi K3 | LMArena Text Elo | 1488 | LMArena text overall, 'Kimi K3 (Max)', ±4, as of 2026-10-27 | Arena (LMArena) text leaderboard |
Scores are published results from model cards, lab reports, and public leaderboards, each linked to its source. Settings such as thinking mode or test version differ between sources and are shown next to each score. OpenLLM Buddy catalog models are uncensored fine-tunes; their scores are those of the official base model. Data collected 2026-10-30. How the index works