Llama · Llama 4 · 400B parameters · 17B active
Llama 4 Maverick 400B-A17B VRAM requirements
Llama 4 Maverick 400B-A17B has 48 layers and uses grouped-query attention (8 KV heads). At Q4_K_M the weights come to 226.0 GB.
Won't fit
228.3 GB of 21.8 GB · 1049%Short by 206.6 GB. You can run it with 4 of 48 layers on the RTX 4090 and the rest in system RAM, at roughly 4.00 tok/s — usable for batch work, painful for chat. A smaller quantisation or a shorter context is usually the better trade.
Every quantisation of Llama 4 Maverick 400B-A17B on a RTX 4090
Highlighted row is the highest quality that still fits at 8K context.
| Quantisation | bpw | Weights | Total | Verdict | Max ctx | tok/s |
|---|---|---|---|---|---|---|
| FP16 / BF16 | 16.00 | 745.1 GB | 747.4 GB | Won't fit | — | 1.20 |
| INT8 / W8A8 | 8.50 | 395.3 GB | 397.7 GB | Won't fit | — | 2.27 |
| Q8_0 (GGUF) | 8.50 | 395.3 GB | 397.7 GB | Won't fit | — | 2.27 |
| FP8 (E4M3) | 8.00 | 372.5 GB | 374.9 GB | Won't fit | — | 2.40 |
| Q6_K | 6.56 | 305.5 GB | 307.8 GB | Won't fit | — | 2.95 |
| Q5_K_M | 5.67 | 264.0 GB | 266.4 GB | Won't fit | — | 3.39 |
| Q5_K_S | 5.52 | 257.0 GB | 259.4 GB | Won't fit | — | 3.47 |
| Q4_K_M | 4.85 | 226.0 GB | 228.3 GB | Won't fit | — | 4.00 |
| Q4_K_S | 4.58 | 213.5 GB | 215.8 GB | Won't fit | — | 4.21 |
| Q4_0 | 4.55 | 212.1 GB | 214.4 GB | Won't fit | — | 4.24 |
| AWQ 4-bit | 4.25 | 200.7 GB | 203.1 GB | Won't fit | — | 4.46 |
| GPTQ 4-bit | 4.25 | 200.7 GB | 203.1 GB | Won't fit | — | 4.46 |
| MXFP4 | 4.25 | 200.7 GB | 203.1 GB | Won't fit | — | 4.46 |
| IQ4_XS | 4.25 | 198.2 GB | 200.5 GB | Won't fit | — | 4.51 |
| Q3_K_M | 3.91 | 182.5 GB | 184.8 GB | Won't fit | — | 4.97 |
| IQ3_M | 3.70 | 172.7 GB | 175.1 GB | Won't fit | — | 5.22 |
| IQ3_XXS | 3.06 | 143.1 GB | 145.4 GB | Won't fit | — | 6.32 |
| Q2_K | 2.63 | 123.2 GB | 125.5 GB | Won't fit | — | 7.37 |
| IQ2_XXS | 2.06 | 96.8 GB | 99.1 GB | Won't fit | — | 9.48 |
| IQ1_M | 1.75 | 82.4 GB | 84.7 GB | Won't fit | — | 11.4 |
Llama 4 Maverick 400B-A17B on each GPU
Q4_K_M weights at 8K context, single card, monitor attached.
| GPU | VRAM | GB/s | Verdict | Max ctx | tok/s |
|---|---|---|---|---|---|
| Mac Studio M3 Ultra 256GB | 256 | 819 | Won't fit | — | 14.6 |
| H100 SXM 80GB | 80 | 3350 | Won't fit | — | 6.20 |
| Mac Studio M4 Max 128GB | 128 | 546 | Won't fit | — | 6.01 |
| NVIDIA DGX Spark (GB10) | 128 | 273 | Won't fit | — | 5.71 |
| A100 80GB | 80 | 2039 | Won't fit | — | 5.51 |
| RTX A6000 | 48 | 768 | Won't fit | — | 4.53 |
| RTX 5090 | 32 | 1792 | Won't fit | — | 4.52 |
| Ryzen AI Max+ 395 128GB | 128 | 256 | Won't fit | — | 4.49 |
| Mac Mini M4 Pro 48GB | 48 | 273 | Won't fit | — | 4.43 |
| L40S | 48 | 864 | Won't fit | — | 4.35 |
| RTX 5080 | 16 | 960 | Won't fit | — | 4.23 |
| RTX 5070 Ti | 16 | 896 | Won't fit | — | 4.23 |
| RTX 5060 Ti 16GB | 16 | 448 | Won't fit | — | 4.21 |
| RTX 3090 | 24 | 936 | Won't fit | — | 4.17 |
| RTX 5070 | 12 | 672 | Won't fit | — | 4.14 |
| RTX 4090 | 24 | 1008 | Won't fit | — | 4.00 |
| RTX 3080 10GB | 10 | 760 | Won't fit | — | 3.92 |
| RTX 3060 12GB | 12 | 360 | Won't fit | — | 3.92 |
| RTX 4080 Super | 16 | 736 | Won't fit | — | 3.83 |
| RTX 4070 Ti Super | 16 | 672 | Won't fit | — | 3.83 |
| RTX 4060 Ti 16GB | 16 | 288 | Won't fit | — | 3.81 |
| Radeon RX 7900 XTX | 24 | 960 | Won't fit | — | 3.79 |
| RTX 4070 Super | 12 | 504 | Won't fit | — | 3.76 |
| RTX 4070 | 12 | 504 | Won't fit | — | 3.76 |
| Arc B580 | 12 | 456 | Won't fit | — | 3.18 |
Architecture
| Parameters | 400B |
| Active per token | 17B of 128 experts, top-1 |
| Layers | 48 |
| Hidden size | 5120 |
| Attention heads / KV heads | 40 / 8 |
| Head dimension | 128 |
| Vocabulary | 202,048 |
| Trained context | 1024K |
| KV cache per 1K tokens | 0 GB |
| Hugging Face | meta-llama/Llama-4-Maverick-17B-128E-Instruct |
The Llama family
Meta's open-weight series, and the default target for most local tooling. Every Llama 3.x model uses grouped-query attention with 8 KV heads, so the cache stays modest even at 70B. Llama 4 moved to mixture-of-experts: Scout and Maverick occupy 109B and 400B of memory but read only 17B per token.
Direct answers
32 GBLlama 4 Maverick 400B-A17B on RTX 4090
24 GBLlama 4 Maverick 400B-A17B on RTX 3090
24 GBLlama 4 Maverick 400B-A17B on RTX 5080
16 GBLlama 4 Maverick 400B-A17B on RTX 5070 Ti
16 GBLlama 4 Maverick 400B-A17B on RTX 5070
12 GBLlama 4 Maverick 400B-A17B on RTX 4070 Ti Super
16 GBLlama 4 Maverick 400B-A17B on RTX 4070 Super
12 GBLlama 4 Maverick 400B-A17B on RTX 4070
12 GB