Llama · Mixture-of-experts · Llama 4 Community
Can I run Llama 4 Scout (MoE) locally?
A 109B MoE with only 17B active and a huge context. Wants ~60 GB+ — a 96–128 GB Mac or multi-GPU box.
- Parameters
- 109B (17B active)
- VRAM at Q4
- 66 GB
- Max context
- 1049K
- Released
- 2025-04
Memory needed by quantisation
Total includes weights, an 8K-token KV cache and runtime overhead. Lower quants trade quality for size.
| Quant | Bits/weight | Weights | Total needed | Quality |
|---|---|---|---|---|
| FP16 | 16 | 203 GB | 213 GB | Full precision. Reference quality, 2× the size of 8-bit. |
| Q8_0 | 8.5 | 108 GB | 114 GB | Effectively lossless. The safe choice when it fits. |
| Q6_K | 6.56 | 83 GB | 89 GB | Near-lossless; quality loss is hard to measure. |
| Q5_K_M | 5.67 | 72 GB | 77 GB | Very good. A common sweet spot above Q4. |
| Q4_K_M | 4.83 | 61 GB | 66 GB | The default. Best size/quality trade-off for local use. |
| Q3_K_M | 3.91 | 50 GB | 54 GB | Noticeable degradation; useful to squeeze a size up. |
| Q2_K | 3.35 | 43 GB | 46 GB | Aggressive. Quality drops a lot — last resort to fit. |
Which hardware runs Llama 4 Scout (MoE)?
Best quantisation that fits each device at 8K context, with a rough speed estimate. Try your exact setup →
| Hardware | Memory | Fit | Best quant | Speed |
|---|---|---|---|---|
| A100 80 GB | 80 GB | Runs well | Q5 | ~79 tok/s |
| H100 80 GB | 80 GB | Runs well | Q5 | ~130 tok/s |
| Mac · M3/M4 Max, 128 GB | 128 GB | Runs well | Q6 | ~18 tok/s |
| Mac Studio · M1/M2 Ultra, 128 GB | 128 GB | Runs well | Q6 | ~27 tok/s |
| Mac Studio · M3 Ultra, 256 GB | 256 GB | Runs well | Q8 | ~21 tok/s |
| Mac Studio · M3 Ultra, 512 GB | 512 GB | Runs well | FP16 | ~11 tok/s |
| CPU only · 128 GB RAM | 128 GB | Runs well | Q8 | ~2.3 tok/s |
| RTX A6000 | 48 GB | Runs (tight) | Q2 | ~50 tok/s |
| RTX 6000 Ada | 48 GB | Runs (tight) | Q2 | ~63 tok/s |
| CPU only · 64 GB RAM | 64 GB | Runs (tight) | Q3 | ~4.5 tok/s |
| RTX 3060 12 GB | 12 GB | Won't fit | — | — |
| RTX 4060 Ti 8 GB | 8 GB | Won't fit | — | — |
| RTX 4060 Ti 16 GB | 16 GB | Won't fit | — | — |
| RTX 3080 10 GB | 10 GB | Won't fit | — | — |
| RTX 5070 | 12 GB | Won't fit | — | — |
| RTX 4070 Super | 12 GB | Won't fit | — | — |
| Radeon RX 7900 XT | 20 GB | Won't fit | — | — |
| RTX 5070 Ti | 16 GB | Won't fit | — | — |
| RTX 4070 Ti Super | 16 GB | Won't fit | — | — |
| RTX 3090 | 24 GB | Won't fit | — | — |
| Radeon RX 7900 XTX | 24 GB | Won't fit | — | — |
| RTX 5080 | 16 GB | Won't fit | — | — |
| RTX 4080 Super | 16 GB | Won't fit | — | — |
| RTX 4090 | 24 GB | Won't fit | — | — |
| RTX 5090 | 32 GB | Won't fit | — | — |
| Mac · M1/M2/M3 (base), 8 GB | 8 GB | Won't fit | — | — |
| Mac · M1/M2/M3 (base), 16 GB | 16 GB | Won't fit | — | — |
| Mac · M4 (base), 24 GB | 24 GB | Won't fit | — | — |
| Mac · M4 Pro, 48 GB | 48 GB | Won't fit | — | — |
| Mac · M1/M2/M3 Max, 32 GB | 32 GB | Won't fit | — | — |
| Mac · M4 Max, 64 GB | 64 GB | Won't fit | — | — |
| Mac · M1/M2/M3 Max, 64 GB | 64 GB | Won't fit | — | — |
| CPU only · 8 GB RAM | 8 GB | Won't fit | — | — |
| CPU only · 16 GB RAM | 16 GB | Won't fit | — | — |
| CPU only · 32 GB RAM | 32 GB | Won't fit | — | — |
GPUs that run Llama 4 Scout (MoE) well
The most affordable cards in our list that run it at a good quantisation.
Hardware links are affiliate links — they don't change the recommendation.
What Llama 4 Scout (MoE) is good for
Related models
FAQ
How much VRAM does Llama 4 Scout (MoE) need?
At Q4_K_M, Llama 4 Scout (MoE) needs about 66 GB including a 8K-token context and overhead (61 GB for the weights alone). Higher quantisation needs more; see the table for every level.
What is the cheapest way to run Llama 4 Scout (MoE)?
The smallest device that runs it well in our list is the A100 80 GB (80 GB). Anything with at least that much memory should handle it at a usable quantisation.
Is Llama 4 Scout (MoE) good for general chat & assistance?
A 109B MoE with only 17B active and a huge context. Wants ~60 GB+ — a 96–128 GB Mac or multi-GPU box.
Estimates — see how we compute these. Memory figures assume an 8K context; long-context use needs more.