Analyzing Kimi-K3 on the LisanBench Evaluation
scaling01 · x · 2026-07-18
Kimi-K3's Performance on LisanBench
The author points out that Kimi-K3 is currently the strongest open-weight model on LisanBench, outperforming Gemini 3.1 Pro (high) overall, though it still falls short of Opus 4.7 xhigh at a comparable token cost.
Key Conclusions
- Ranks 5th on standard metrics, and 4th on difficulty-weighted metrics and best-of-3 metrics.
- It also outperforms GPT-5.5 and Opus 4.8, but at the cost of using 2–3x the tokens.
- The author believes this aligns with trends observed on other benchmarks: Kimi-K3 is very strong, but still needs a couple of iterations to improve inference efficiency to truly catch up with frontier models.
Additional Observations
- The model's search behavior is noticeably different from other frontier models.
- It uses some very obscure or outdated words, which the author feels puts it at a disadvantage in standard dictionary scoring.
- More notably, it heavily exploits "wildcard families": by repeatedly changing a single letter at the same position, it continuously jumps within a cluster of words with an edit-distance of 1.
- The author states that about 35% of Kimi-K3's valid transitions occur within these families, compared to less than 2% for all other models.
Related event: Kimi-K3 Tops LisanBench as Strongest Open-Weight Model(2 posts)→
More from Models
- BullshitBench update: GPT-6-Astra beats all prior OpenAI models but still trails Anthropic — scaling01 · 2026-09-11
- Astra Scores 83% on GauntletBench, First Computer-Use Agent to Beat Human Baseline — ducha_aiki · 2026-09-11
- Kimi K2.8 Preview rolls out: near-K3 coding performance, 1M context for all tiers — teortaxesTex · 2026-09-11
- Looking for a classifier of software engineering task shapes to pick models per task — StewartalsopIII · 2026-09-11
- DeepSeek V4 Pro API to continue after Sept 2026, billing unchanged — teortaxesTex · 2026-09-11
- DeepSeek V4.1 Flash Hits 98% of GPT-6 Astra's Score at 1.4% of the Cost in Third-Party Benchmark — ayushtweetshere · 2026-09-11