Llama 4 Scout
Llama · Llama 4 Community License
Mixture of Experts: 109B total parameters, 17B active per token.
Large MoE model (~109B total, ~17B active per token; 16 experts, 2 active). Multimodal (text and image) with an enormous advertised context window (10M tokens); practical local runs are usually capped by VRAM and tooling far below that. MoE loads the full weight set for common local inference paths. Not a consumer GPU model: even Q4_K_M needs ~60GB VRAM. Expect H100-class or dual A100 hardware for practical deployment. Do not recommend to RTX 4090 or Mac users.
- Parameters
- 109B
- Architecture
- MoE (17B active)
- Context
- 10,000,000 tokens
- Released
- 2025-04-05
- Engines
- llama.cpp, ollama, vLLM
- Builder Tools
- Continue, LM Studio, Open WebUI
Parameters
109B
VRAM
75 GB
Context
9766K
Formats
4
GPUs
13
Llama 4 Scout (109B) requires 75 GB VRAM at recommended quality (Q5_K_M). On NVIDIA RTX PRO 6000 Blackwell, expect approximately 95 tok/s at Q5_K_M. For the best experience, Mac Studio AI Builder ($3,999) is recommended.
Source: OwnRig methodology
75 GB
Q5_K_M
72 GB
9766K tokens
Chat
VRAM Requirements
| Quality | Quantization | VRAM | File Size |
|---|---|---|---|
| full | Q8_0 | 110 GB | 109 GB |
| recommended | Q5_K_M | 75 GB | 72 GB |
| efficient | Q4_K_M | 60 GB | 58 GB |
| compressed | Q3_K_M | 50 GB | 48 GB |
Context Length Impact
KV cache VRAM at Q5_K_M quality. Longer context = more memory.
| Context | KV Cache | Total VRAM |
|---|---|---|
| 2K | 307 MB | 75.3 GBexceeds 24 GB |
| 4K | 614 MB | 75.6 GBexceeds 24 GB |
| 8K | 1.2 GB | 76.2 GBexceeds 24 GB |
| 16K | 2.4 GB | 77.4 GBexceeds 24 GB |
| 32K | 4.8 GB | 79.8 GBexceeds 24 GB |
| 64K | 9.6 GB | 84.6 GBexceeds 24 GB |
| 128K | 19.2 GB | 94.2 GBexceeds 24 GB |
Compatible GPUs
13 devices| NVIDIA Grace Blackwell Ultra GB300 | Q8_0 | 40 tok/s | Excellent |
| NVIDIA RTX PRO 6000 Blackwell | Q5_K_M | 95 tok/s | Excellent |
| NVIDIA RTX PRO 6000 Blackwell Max-Q | Q5_K_M | 87 tok/s | Excellent |
| Apple M4 Max (128GB Unified) | Q8_0 | 4 tok/s | Marginal |
| Apple M4 Max (64GB Unified) | Q4_K_M | 5 tok/s | Marginal |
| Apple M4 Pro (48GB) | Q3_K_M | 6 tok/s | Marginal |
| Apple M4 Ultra (192GB) | Q8_0 | 5 tok/s | Marginal |
| AMD Radeon Pro W7900 | Q3_K_M | 2 tok/s | Marginal |
| AMD Radeon RX 7600 | Q3_K_M | – | Not viable |
| AMD Radeon RX 7900 XTX | Q3_K_M | – | Not viable |
| AMD Radeon RX 9070 | Q3_K_M | – | Not viable |
| AMD Radeon RX 9060 XT 16GB | Q3_K_M | – | Not viable |
| AMD Radeon RX 9060 XT 8GB | Q3_K_M | – | Not viable |
Showing 13 of 13 entries
Builder Context
Llama 4 Scout is commonly used with Continue, LM Studio, Open WebUI.
Frequently Asked Questions
- How much VRAM does Llama 4 Scout need?
- Llama 4 Scout requires 75 GB VRAM at recommended quality (Q5_K_M). At lower quality settings, it can fit in as little as 50 GB.
- What is the best GPU for Llama 4 Scout?
- The NVIDIA RTX PRO 6000 Blackwell delivers the best performance for Llama 4 Scout, achieving 95 tok/s at Q5_K_M with an excellent rating.
- What quantization should I use for Llama 4 Scout?
- For the best quality, use Q5_K_M (75 GB VRAM). If your GPU has limited VRAM, Q3_K_M (50 GB) is the most efficient option with acceptable quality.
Data confidence: estimated. Source
VRAM requirements are calculated from model parameters and may vary by inference engine, context length, and batch size. Performance estimates are based on community benchmarks and should be verified for your specific configuration.Llama is a trademark of its respective owner. OwnRig is not affiliated with or endorsed by the model creator.