Kimi K3
2.8T-parameter native multimodal agentic model with 1M context. Top of open-weight leaderboards; data-center hardware only (smallest Unsloth GGUF is 594 GB).
Recommended quants
- Unsloth UD-IQ1_S (594 GB)
unsloth/Kimi-K3-GGUF - UD-Q2_K_XL (861 GB)
unsloth/Kimi-K3-GGUF - UD-Q4_K_XL (1.5 TB)
unsloth/Kimi-K3-GGUF
Chat template Caveats
Thinking always on; effort via top-level reasoning_effort (low/high/max). Trained in preserved-thinking mode.
Fix: For multi-turn and tool calls, pass back the COMPLETE assistant message (including reasoning_content). Dropping it degrades results.
Sampling settings
| Moonshot eval | top_p 0.95 (reasoning/vision), 1.0 (agentic); temperature unverified |
|---|
Runtimes
| vLLM | Unverified | |
|---|---|---|
| llama.cpp | Works | Unsloth GGUFs. |
| Ollama | Unverified | Library page kimi-k3 exists; likely cloud. |
| LM Studio | Unverified | |
| MLX | Unverified |
Tool calling in agent tools
| Claude Code (local endpoint) | Unverified | |
|---|---|---|
| OpenCode | Unverified | |
| Cline | Unverified | |
| Aider | Unverified |
Known issues
Sources checked
Spotted something wrong or a new fix? Corrections are reviewed before publishing. (Submission form coming soon.)