Command · released 2026-05 · 219B parameters (25.0B active per token)
Sparse MoE (128 experts top-8 + 4 shared) with 3 sliding (4096) : 1 global layers; 128 query heads x 128 (q dim 16384 > hidden). SigLIP vision encoder. Card: 218B total / 25B active, 128K context (config max_position_embeddings 200000). FP8 and W4A4 checkpoints published.
Source configMemory for one conversation's context, on top of the weights.
| Context | FP16 | FP8 |
|---|---|---|
| 4,096 tokens | 0.50 GB | 0.25 GB |
| 8,192 tokens | 0.63 GB | 0.31 GB |
| 32,768 tokens | 1.38 GB | 0.69 GB |
| 131,072 tokens | 4.38 GB | 2.19 GB |
| Format | Bits | Weights |
|---|---|---|
| BF16 / FP16 | 16 | 407.5 GB |
| FP8 (W8A8) | 8 | 204.7 GB |
| INT8 (W8A8 / W8A16) | 8 | 204.7 GB |
| NVFP4 | 4.5 | 116.0 GB |
| INT4 AWQ / GPTQ | 4.25 | 109.7 GB |
| GGUF Q8_0 | 8.5 | 216.5 GB |
| GGUF Q6_K | 6.56 | 167.1 GB |
| GGUF Q5_K_M | 5.69 | 144.9 GB |
| GGUF Q4_K_M | 4.89 | 124.5 GB |
| GGUF Q3_K_M | 3.91 | 99.6 GB |
| GGUF Q2_K | 3.35 | 85.3 GB |
One server, 8K context, FP8 on data center GPUs and GGUF Q4_K_M on consumer and Apple hardware.
| GPU | 1 user | 32 users |
|---|---|---|
| B200 SXM (HGX B200)FP8 (W8A8) | 2× · 210 tok/s | 2× · 39.9 tok/s |
| H200 SXM 141GBFP8 (W8A8) | 2× · 148 tok/s | 2× · 24.7 tok/s |
| H100 SXM 80GBFP8 (W8A8) | 4× · 151 tok/s | 4× · 32.4 tok/s |
| Instinct MI300X 192GBFP8 (W8A8) | 2× · 159 tok/s | 2× · 27.1 tok/s |
| A100 SXM 80GBINT4 AWQ / GPTQ | 2× · 124 tok/s | 2× · 18.9 tok/s |
| L40S 48GBFP8 (W8A8) | 8× · 53.2 tok/s | 8× · 14.9 tok/s |
| RTX PRO 6000 Blackwell Server Edition 96GBFP8 (W8A8) | 4× · 75.8 tok/s | 4× · 15.5 tok/s |
| GeForce RTX 5090 32GBGGUF Q4_K_M | 8× · 64.7 tok/s | 8× · 31.3 tok/s |
| GeForce RTX 4090 24GBGGUF Q4_K_M | 8× · 58.1 tok/s | 8× · 21.4 tok/s |
| Apple M4 Max (40-core GPU, up to 128GB)GGUF Q4_K_M | > 16 | > 16 |
| DGX Spark (GB10, 128GB unified)FP8 (W8A8) | > 16 | > 16 |