Llama 4 Scout VRAM Requirements
Llama 4 Scout needs about 64.0 GB of VRAM at Q4_K_M with an 8K context: 61.5 GB for weights, 1.5 GB for the KV cache and 1.0 GB of runtime overhead. At Q8_0 it needs 110 GB and at full FP16 precision 205 GB. The smallest GPU that runs it fully is the A100 80GB.
64.0 GB VRAM
Llama 4 Scout at Q4_K_M with 8K context and the KV cache in FP16.
- Weights 61.5 GB
- KV cache 1.5 GB
- Runtime overhead 1.0 GB
Mixture of experts: all 109B parameters need memory, but only 17B are read per token, so it runs at the speed of a much smaller model.
Which GPUs can run it
- RTX 3060 12GBToo small
- RTX 4060 Ti 16GBToo small
- RTX 3090 24GB35% on GPU~5.5 tok/s
- Radeon RX 7900 XTX 24GB35% on GPU~5.4 tok/s
- RTX 4090 24GB35% on GPU~5.5 tok/s
- RTX 5090 32GB47% on GPU~6.7 tok/s
- RTX 6000 Ada 48GB70% on GPU~10 tok/s
- A100 80GBFits~76 tok/s
- H100 80GB SXMFits~101 tok/s
- Ryzen AI Max+ 395 128GBFits~13 tok/s
- Apple M4 Max 128GBFits~28 tok/s
- Apple M5 Max 128GBFits~31 tok/s
- RTX PRO 6000 Blackwell 96GBFits~70 tok/s
- NVIDIA DGX Spark 128GBFits~14 tok/s
- Apple M3 Ultra 512GBFits~39 tok/s
Multi-GPU: about 3× 24 GB cards (RTX 3090 / 4090) with the model split across them.
Llama 4 Scout VRAM by quantization
Total memory with an FP16 KV cache. "Smallest tier" is the smallest standard VRAM size with a 6% safety margin at 8K.
| Quantization | Bits/weight | Weights | Total at 8K | Total at 32K | Smallest tier |
|---|---|---|---|---|---|
| Q4_K_M | 4.85 | 61.5 GB | 64.0 GB | 65.1 GB | 80 GB |
| Q5_K_M | 5.69 | 72.2 GB | 74.7 GB | 75.8 GB | 80 GB |
| Q6_K | 6.59 | 83.6 GB | 86.1 GB | 87.2 GB | 96 GB |
| Q8_0 | 8.50 | 108 GB | 110 GB | 111 GB | 128 GB |
| FP16 / BF16 | 16.00 | 203 GB | 205 GB | 207 GB | Multi-GPU / server |
| Q3_K_M | 3.91 | 49.6 GB | 52.1 GB | 53.2 GB | 64 GB |
| Q2_K | 3.00 | 38.1 GB | 40.5 GB | 41.7 GB | 48 GB |
VRAM by context length
At Q4_K_M. 12 of 48 layers keep the full context. The other 36 use a 8192-token sliding window, so their cache stops growing after 8K tokens.
| Context | KV cache (FP16) | KV cache (Q8_0) | Total, FP16 cache | Total, Q8_0 cache |
|---|---|---|---|---|
| 4K tokens | 0.8 GB | 0.4 GB | 63.3 GB | 62.9 GB |
| 8K tokens | 1.5 GB | 0.8 GB | 64.0 GB | 63.3 GB |
| 32K tokens | 2.6 GB | 1.4 GB | 65.1 GB | 63.9 GB |
| 128K tokens | 7.1 GB | 3.8 GB | 69.6 GB | 66.3 GB |
GPUs that can run Llama 4 Scout
At Q4_K_M with 8K context. "Best quant" is the highest-quality format that fits fully; "max context" is at Q4_K_M.
| GPU | Usable memory | Status | Best quant | Max context | Speed |
|---|---|---|---|---|---|
| A100 80GB | 80.0 GB | Fits | Q5_K_M | 128K | ~76 tok/s |
| H100 80GB SXM | 80.0 GB | Fits | Q5_K_M | 128K | ~101 tok/s |
| Ryzen AI Max+ 395 128GB | 96.0 GB | Fits | Q6_K | 128K | ~13 tok/s |
| Apple M3 Max 128GB | 96.0 GB | Fits | Q6_K | 128K | ~21 tok/s |
| Apple M4 Max 128GB | 96.0 GB | Fits | Q6_K | 128K | ~28 tok/s |
| Apple M5 Max 128GB | 96.0 GB | Fits | Q6_K | 128K | ~31 tok/s |
| RTX PRO 6000 Blackwell 96GB | 96.0 GB | Fits | Q6_K | 128K | ~70 tok/s |
| NVIDIA DGX Spark 128GB | 112 GB | Fits | Q6_K | 128K | ~14 tok/s |
| NVIDIA H200 141GB | 141 GB | Fits | Q8_0 | 128K | ~120 tok/s |
| Apple M2 Ultra 192GB | 144 GB | Fits | Q8_0 | 128K | ~38 tok/s |
| AMD Instinct MI300X 192GB | 192 GB | Fits | Q8_0 | 128K | ~116 tok/s |
| Apple M3 Ultra 512GB | 384 GB | Fits | FP16 / BF16 | 128K | ~39 tok/s |
| RTX 6000 Ada 48GB | 48.0 GB | 70% on GPU | Q2_K | – | ~10 tok/s |
| RTX A6000 48GB | 48.0 GB | 70% on GPU | Q2_K | – | ~10 tok/s |
| L40S 48GB | 48.0 GB | 70% on GPU | Q2_K | – | ~10 tok/s |
| A100 40GB | 40.0 GB | 59% on GPU | – | – | ~8.3 tok/s |
| RTX 5090 32GB | 32.0 GB | 47% on GPU | – | – | ~6.7 tok/s |
| Radeon AI PRO R9700 32GB | 32.0 GB | 47% on GPU | – | – | ~6.3 tok/s |
Multi-GPU: about 3× 24 GB cards (RTX 3090 or 4090) run it at Q4_K_M, using tensor or layer splitting in llama.cpp, vLLM or ExLlamaV2.
About Llama 4 Scout
A mixture-of-experts model with 16 experts: 109B total parameters, 17B active per token. All 109B must sit in memory, but speed behaves like a 17B model. Three of every four layers use chunked attention over 8K tokens, so the KV cache grows slowly with context. Meta advertises a 10M-token window; this calculator caps it at 128K.
| Spec | Value |
|---|---|
| Parameters | 109B total, 17B active (MoE) |
| Layers | 48 |
| Hidden size | 5120 |
| Attention heads / KV heads | 40 / 8 |
| Head dimension | 128 |
| Vocabulary | 202,048 |
| Max context (used here) | 128K tokens |
| Sliding window | 8192 tokens on 36 of 48 layers |
| KV cache per token (FP16) | 48 KB |
| Released | 2025-04 |
| License | Llama 4 Community License |
Llama 4 Scout VRAM: common questions
Does Llama 4 Scout need VRAM for 17B or 109B parameters?
For 109B. Every expert must be loaded because any token can route to any expert. The 17B active figure only sets the speed.
How much VRAM does Llama 4 Scout need?
About 64.0 GB at Q4_K_M with an 8K context, 110 GB at Q8_0 and 205 GB at FP16. Longer contexts add 0.05 GB per 1,000 tokens with an FP16 cache.
What is the cheapest GPU that can run Llama 4 Scout?
The smallest listed device that fits it fully at Q4_K_M with 8K context is the A100 80GB (80.0 GB usable). With 24 GB cards you need about 3 of them.
Does Llama 4 Scout need VRAM for all 109B parameters?
Yes. All experts must be loaded because any token can use any expert. Only 17B parameters are read per token, so it generates text about as fast as a 17B dense model.