Kimi Linear 48B-A3B
by Moonshot AI
· Kimi Linear family · released Oct 30, 2025
Open weights
Active
Moonshot's 48B hybrid linear-attention model combining Kimi Delta Attention with gated MoE (3B active), cutting KV cache size by up to 75% while outperforming full-attention baselines in long-context tasks.
This model has no first-party API. Open-weight release: self-host from published weights.
Specifications
| Model type | Text LLM |
|---|---|
| Architecture | Hybrid |
| Parameters | 48 (3 active) |
| Context window | Not disclosed tokens |
| Max output | Not disclosed tokens |
| Knowledge cutoff | Not disclosed |
| License | Apache License 2.0 |
| Modalities | Text input Text output |
| Tags | Long context |
| End of life | Not disclosed |
| Last verified | Not disclosed |