// Kimi · Vision-language
Kimi-K3: VRAM & local deployment
2.8T total · 104B active (MoE) parameters. About 1.87 TB at 4-bit and 6.16 TB at BF16 with one 8K-token request.
released 2026-06-13licence: Custom licence – read itmemory figures are estimates
01Specs
| Publisher | Moonshot AI · moonshotai |
|---|---|
| Family / series | Kimi · Kimi-K3 |
| Type | Vision-language (image-text-to-text) |
| Parameters | 2.8T total · 104B active (MoE) Source: official model card“Total Parameters 2.8T · Activated Parameters 104B” |
| Context length | 1M tokens (1,048,576)max_position_embeddings in config.json |
| Architecture | 93 layers (24 full-attention) · MLA+HYBRID · MoE |
| Released weights | packed quantised tensors (U8, e.g. MXFP4 / INT4) · 1.56 TBdownload |
| Licence | kimi-k3 · Custom licence – read itThe publisher uses its own licence; we have not classified it. Read the licence text. |
| Released | 2026-06-13 (Hugging Face repo created) |
02How much VRAM it needs (estimate)
Total = weights + 10% + 0.5GB overhead + KV cache for one request (FP16 KV). More concurrent users multiply the KV column.
| Precision | Weights | Total @ 8K | Total @ 32K | Total @ 128K | Smallest GPU | Mac |
|---|---|---|---|---|---|---|
| BF16 / FP16 (original) | 5.60 TB | 6.16 TB | 6.16 TB | 6.16 TB | multi-node | — |
| FP8 / INT8 / Q8_0 | 2.98 TB | 3.27 TB | 3.27 TB | 3.28 TB | 2 × 8 × B300 nodes (4,608GB) | — |
| 4-bit (Q4_K_M / AWQ / GPTQ-Int4) | 1.70 TB | 1.87 TB | 1.87 TB | 1.87 TB | 8 × B300 (2,304GB) | — |
KV cache per request: 8K 0.2 GB · 32K 0.9 GB · 128K 3.6 GB
Bits per weight: BF16 16, FP8 8.5, 4-bit 4.85. “Fits” = total ≤ 90% of memory; Mac GPU can use ~2/3 of unified memory (≤36GB) or ~3/4 (larger). For MoE, all experts must be in memory, so size by total parameters. Method →
03How to run it
Only runtimes the publisher or the runtime’s official library names for this model.
- vLLM — mentioned in the official model card
- SGLang — mentioned in the official model card
- Transformers — mentioned in the official model card
04Official repositories & quantised builds
1 official repo(s) from moonshotai.
- Kimi-K3 main
05Other Kimi models
All 14 →| Model | Params | 4-bit total | Released |
|---|---|---|---|
| Kimi-K2.7-Code | 1T / 32B | 668 GB | 2026-06-11 |
| Kimi-K2.6 | 1T / 32B | 668 GB | 2026-04-14 |
| Kimi-K2.5 | 1T / 32B | 668 GB | 2026-01-01 |
| Kimi-K2-Thinking | 1T / 32B | 668 GB | 2025-11-04 |
| Kimi-Linear-48B-A3B-Instruct | 49.1B / 3B | 33.3 GB | 2025-10-30 |
| Kimi-K2-Instruct-0905 | 1T / 32B | 668 GB | 2025-09-03 |
| Kimi-K2-Instruct | 1T / 32B | 668 GB | 2025-07-11 |
| Kimi-VL-A3B-Thinking-2506 | 16.4B / 3B | 11.7 GB | 2025-06-21 |
06Sources
Data checked 2026-09-02. Parameters, context and licence are copied from the official repository; memory figures are our estimates.