Kimi K3 2.8T-A50B: the largest open model ever released; Opus 4.8-class at Sonnet 5 pricing
- Source
- latent.space
- Author
- Latent Space
- Date
Context
Kimi K3's surrounding software needs to preserve its earlier reasoning during an ongoing task. Moonshot AI says the model was trained with that history retained, and warns that generation can become unstable when an framework leaves it out or switches an existing conversation from another model to K3. The requirement appears in its technical blog.
For developers evaluating K3 for extended coding work, compatibility therefore includes how the application carries a session forward. Moonshot recommends a verified compatible framework and avoiding a model switch midway through a session. A successful initial response alone would not establish that the integration handles the history K3 expects.
- AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
- token — The chunk of text a model reads and writes in — roughly three-quarters of a word — and the unit AI usage is billed in.
- context window — The maximum amount of text a model can consider at once — its working memory for the current conversation or task.
- multimodal — A model that works with more than text — reading images, audio, or video, and sometimes generating them too.
“Arena later added that K3 has a 76% pairwise win rate in Frontend Code Arena, versus 63% for Fable 5 and 58% for GPT-5.6 Sol”
“Artificial Analysis published an independent evaluation placing K3 at 57 on the AA Intelligence Index , calling it comparable to Opus 4.8 and GPT-5.5 , but still behind Fable 5 and GPT-5.6 Sol overall”
“K3 uses Kimi Delta Attention (KDA) , which Moonshot says enables up to 6.3x faster decoding in million-token contexts”
“The strongest repeated theme: frontier-ish performance at materially lower price than top closed models , though not at bargain-basement open-model prices”
Checking sign-in…
Loading comments…