Kimi K3: Moonshot AI's 2.8T Open Model Claims Third Place Behind Claude Fable 5 and GPT 5.6 Sol
- Kimi K3 is a 2.8 trillion parameter model with a 1 million token context window, native vision, and is described as the world's first open 3T class model, built on new Kimi Delta Attention (KDA) and Attention Residuals (AttnRes) architecture.
- On Moonshot's own evaluation suite K3 trails only Claude Fable 5 and GPT 5.6 Sol among tested models, though it does not surpass either.
- The Stable LatentMoE framework activates just 16 of 896 experts, and combined with training and data improvements gives roughly a 2.5x gain in scaling efficiency over the prior Kimi K2 model.
- API pricing is $3 per million input tokens and $15 per million output tokens (cache at $0.3), matching Anthropic's Sonnet pricing and drawing criticism on Hacker News as expensive for a Chinese open weight model.
- Full model weights are set for release by July 27, 2026, with a technical report to follow covering architecture, training, and evaluation details; the model is already live on Kimi.com, Kimi Work, Kimi Code, and the Kimi API.
Hacker News 의견들
1M context and pricing is $3/$15 per million tokens with $0.3 cache, which is really high for a Chinese open weight model. But if it's actually close to Fable/Sol level, that pricing might be justified, it's basically 1:1 with Anthropic's Sonnet pricing.
One thing people forget: reasoning efficiency matters a ton for real cost. GPT models are super token efficient, so if Sol uses 10K reasoning tokens and K3 needs 50K to do the same task, Sol still wins on total cost even at a higher per token price.
The quickstart page is weirdly thin, it just points to the tech blog for benchmarks but K3 isn't even mentioned there yet, last post is still about K2.6.
Tokenizers matter too, Anthropic's encodes the same text at way higher token counts than OpenAI's. Also I don't think Kimi is really competing with Fable/Sol, it's competing with GLM, and GLM 5.2 is less than a third of the price.
I burn through 1M context in a local model in 3 to 4 hours of normal use. It'd have to be extremely smart and error free for that context window to actually matter in practice.
Kind of wild that with the Stable LatentMoE setup activating 16 of 896 experts, if you do the naive math that's around 50B active params out of 2.8T total. Not sure that math is right though.
No, you can't just divide total size by expert count like that, a lot of weights are shared/constant across all tokens so the real active param formula accounts for that shared portion separately.
2.5x scaling efficiency over K2 but the price is like 4x higher, so what's actually happening there, did the subsidy era just end for Chinese labs?
I've been using Fable a lot and it's great but its reasoning trace is opaque. Kimi actually exposes the full reasoning trace, which makes debugging its logic way more informative and honestly kind of fun to watch.
Cool to see LatentMoE getting picked up by more labs. Also this kind of implies Fable and Sol themselves are probably in the 4 to 5T parameter range if K3 at 2.8T is only slightly behind them.
Inference on a 2.8T model is genuinely expensive even with margins baked in, if they release the open weights properly the price could drop to like $10 to $12 per million tokens eventually.
Excited for the DeepSeek release that's supposedly coming this week too, hoping they push closer to SOTA as well.
DeepSeek's cache pricing is absurdly cheap, something like $0.003 per million tokens, like 100x cheaper than Kimi K3's cache price. I use it constantly for long agent sessions because it costs basically nothing.