How Much VRAM Does Llama 4 Scout 17B 16E Need to Run Locally?
See devices that run itLlama 4 Scout 17B 16E is a MoE model with 108.6B total parameters and 17B active per token. At Q4 it needs 64GB of VRAM to sit entirely on the GPU; with experts offloaded to system RAM it needs only 10GB of VRAM plus 55GB of RAM. The cheapest device that holds it is the CMP 170HX 64GB (modded), with no used-price data yet, at roughly 71.5 tokens/s. In offload mode the cheapest card that runs it is the GeForce RTX 2080 Ti, at roughly 11 tokens/s.
Llama 4 Scout 17B 16E details
VRAM needed for Llama 4 Scout 17B 16E by quantization and context
Total at 8KMoE: all 108.6B parameters must stay resident in memory, while speed is set by the 17B active per token.
| Quantization | WeightsGB | Total at 8KGB | Total at 32KGB | Total at 128KGB |
|---|---|---|---|---|
| Q4_K_M | 65.4GGUF | 63.1 | 67.6 | 85.6 |
| Q5_K_M | 76.5GGUF | 75.8 | 80.3 | 98.3 |
| Q8_0 | 112.6est. | 115.1 | 119.6 | 137.6 |
| FP16 | 215.6GGUF | 215 | 219.5 | 237.5 |
Total = weights + KV cache (FP16) + 1GB runtime overhead. Weights marked GGUF are measured file sizes from the repository; the rest are estimated from bytes per parameter. Contexts beyond this model's 10M limit show a dash.
Which GPUs can run Llama 4 Scout 17B 16E?
Q4_K_M · 8K context · sorted by eBay used price| Device | VRAMGB | Used price | Est. t/s | Suggested context |
|---|---|---|---|---|
 CMP 170HX 64GB (modded) | 64 | —No listings | 71.5 | 8K |
 Instinct MI210 | 64 | —No listings | 59.7 | 8K |
| RTX PRO 5000 Blackwell 72GB | 72 | —No listings | 66.2 | 32K |
| RTX PRO 6000D | 84 | —No listings | 74 | 32K |
| RTX PRO 6000 Blackwell | 96 | —No listings | 81.2 | 128K |
| Mac Studio M4 Max 128GB | 128 | —No listings | 25.8 | 128K |
| Mac Studio M5 Max 128GB | 128 | —No listings | 28.6 | 128K |
| MacBook Pro M4 Max 128GB | 128 | —No listings | 25.8 | 128K |
| MacBook Pro M5 Max 128GB | 128 | —No listings | 28.6 | 128K |
| Mac Studio M3 Ultra 512GB | 512 | —No listings | 36.8 | 128K |
Fits when weights + KV cache + runtime overhead is at or below usable memory, single card. "Offload" on MoE models means it runs with expert weights in system RAM (only attention layers and KV cache stay in VRAM); that speed assumes 70 GB/s RAM bandwidth. The context column is the largest tier that fits.
GPUs that run Llama 4 Scout 17B 16E with experts offloaded to RAM
Expert weights live in system RAM (needs at least 55 GB); VRAM holds only attention layers, shared experts and the KV cache, at least 10 GB at Q4. Speed is bound by RAM bandwidth (estimated at 70 GB/s) and is far slower than a full-VRAM setup.
| Device | VRAMGB | Used price | RAM neededGB | Est. t/s |
|---|---|---|---|---|
| GeForce RTX 2080 Ti | 11 | — | 55 | 11 |
| GeForce RTX 4080 | 16 | — | 55 | 11.2 |
| GeForce RTX 5070 Ti | 16 | — | 55 | 11.6 |
| GeForce RTX 5080 | 16 | — | 55 | 11.7 |
| Radeon RX 7800 XT | 16 | — | 55 | 10.3 |
| Radeon RX 9070 XT | 16 | — | 55 | 10.3 |
| Tesla V100 16GB | 16 | — | 55 | 11.6 |
| Radeon RX 7900 XT | 20 | — | 55 | 10.8 |
 Arc Pro B60 | 24 | — | 55 | 9.2 |
| GeForce RTX 3090 | 24 | — | 55 | 11.7 |
| GeForce RTX 4090 | 24 | — | 55 | 11.8 |
| GeForce RTX 5090 D V2 | 24 | — | 55 | 12.1 |
| Radeon RX 7900 XTX | 24 | — | 55 | 11.2 |
| RTX PRO 4000 Blackwell | 24 | — | 55 | 11.1 |
| Arc Pro B65 | 32 | — | 55 | 10 |
| Arc Pro B70 | 32 | — | 55 | 10 |
| GeForce RTX 4080 SUPER 32GB (modded) | 32 | — | 55 | 11.3 |
| GeForce RTX 5090 | 32 | — | 55 | 12.4 |
| GeForce RTX 5090 D | 32 | — | 55 | 12.4 |
| Instinct MI100 | 32 | — | 55 | 11.6 |
| Radeon AI PRO R9700 | 32 | — | 55 | 10.3 |
| Radeon PRO W7800 | 32 | — | 55 | 10.1 |
| RTX PRO 4500 Blackwell | 32 | — | 55 | 11.6 |
| Tesla V100 32GB | 32 | — | 55 | 11.6 |
| A100 40GB PCIe | 40 | — | 55 | 12.3 |
| CMP 170HX 40GB (modded) | 40 | — | 55 | 12.3 |
| GeForce RTX 4090 48GB (modded) | 48 | — | 55 | 11.8 |
| Radeon PRO W7900 | 48 | — | 55 | 11 |
| RTX PRO 5000 Blackwell 48GB | 48 | — | 55 | 12.1 |
Run Llama 4 Scout 17B 16E with Ollama, llama.cpp or vLLM
Ollama pulls and runs it in one command; the quantization comes from the official tag.
ollama run llama4:scout
llama.cpp pulls the GGUF straight from Hugging Face; -c sets the context length.
llama-server -hf unsloth/Llama-4-Scout-17B-16E-Instruct-GGUF:Q4_K_M -c 8192
vLLM serves the original-precision weights, which needs far more memory than GGUF and usually more than one GPU.
vllm serve meta-llama/Llama-4-Scout-17B-16E-Instruct
File names follow that week's model-library snapshot; the definitions are in the methodology. Methodology
FAQ
How much VRAM does Llama 4 Scout 17B 16E need?
At Q4_K_M with an 8K context Llama 4 Scout 17B 16E needs about 64GB of VRAM; 76GB leaves comfortable headroom.
What is the cheapest GPU that runs Llama 4 Scout 17B 16E?
The CMP 170HX 64GB (modded): 64GB of VRAM at roughly 71.5 tokens/s. We have no used-price data for it yet.
Can you run Llama 4 Scout 17B 16E with Ollama?
Yes: ollama run llama4:scout.