Llama 4 Scout 17B 16E needs ~83.7 GB VRAM. NVIDIA DGX Spark 128GB has 108.8 GB. With Q4_K_M quantization, expect ~6 tok/s.
Operating mode
Interactive favors responsiveness, while light API and scale-out lean harder on serving readiness. The fit stays the same, but the recommendation lens changes.
Current mode
Balanced
Balanced for general local use. Keeps the ranking neutral across personal and serving workflows.
Select quantization to explore
Fit status
Runs well
Decode
6.3 tok/s
TTFT
30909 ms
Safe context
153K
Memory
83.7 GB / 108.8 GB
The model fits in shared memory, but shared-memory bandwidth is now the real limiter.
Fit does not mean dedicated-VRAM speed
Unified or shared memory can make a model technically fit, but sustained tokens per second may still trail a discrete high-bandwidth GPU with less total memory.
Shared-memory contention still exists
The OS, browser, and inference runtime all compete for the same physical memory pool, so real-world headroom is less forgiving than raw capacity suggests.
Prioritize bandwidth, not only capacity
If this workload feels slow, the next useful step is often a GPU tier with materially faster memory bandwidth rather than only a small bump in capacity.
| Workload | Grade | Fit | Decode | TTFT | Context |
|---|---|---|---|---|---|
| Chat | A | Runs well | 6.3 tok/s | 16859 ms | 153K |
| Coding | A | Runs well | 6.3 tok/s | 30909 ms | 153K |
| Agentic Coding | A | Runs well | 6.3 tok/s | 44958 ms | 153K |
| Reasoning | A | Runs well | 6.3 tok/s | 36528 ms | 153K |
| RAG | A | Runs well | 6.3 tok/s | 56197 ms | 153K |
Inference speed
Estimated decode speed (tokens/sec) for Llama 4 Scout 17B 16E at Q4_K_M across popular GPUs and Apple Silicon, including multi-GPU rigs, using the fastest local runtime per device. Fastest is Mac Studio M3 Ultra 256GB at ~21 tok/s. Speed is memory-bandwidth bound, so cards that fit the whole model in VRAM run far faster than ones that offload to system RAM.
| GPU / Mac | Memory | Quant | Speed (tok/s) | Fits? |
|---|---|---|---|---|
Mac Studio M3 Ultra 256GB | 256 GB | Q4_K_M | 21.3 | Fits |
Mac Studio M2 Ultra 128GB | 128 GB | Q4_K_M | 17.7 | Tight |
Mac Studio M1 Ultra 128GB | 128 GB | Q4_K_M | 16.8 | Tight |
MacBook Pro M4 Max 128GB | 128 GB | Q4_K_M | 13.2 | Tight |
2× RX 7900 XTX 24GB | 48 GB | Q4_K_M | 13.1 | Too big |
 2× RTX 4090 24GB | 48 GB | Q4_K_M | 8.7 | Too big |
| 2× RTX 3090 24GB | 48 GB | Q4_K_M | 7.5 | Too big |
| RTX 5090 32GB | 32 GB | Q4_K_M | 6.9 | Too big |
MacBook Pro M4 Max 64GB | 64 GB | Q4_K_M | 6.8 | Too big |
| 4× RTX 3060 12GB | 48 GB | Q4_K_M | 6.6 | Too big |
MacBook Pro M3 Max 64GB | 64 GB | Q4_K_M | 4.7 | Too big |
| RTX 4090 24GB | 24 GB | Q4_K_M | 4.4 | Too big |
MacBook Pro M1 Max 64GB | 64 GB | Q4_K_M | 4.3 | Too big |
RX 7900 XTX 24GB | 24 GB | Q4_K_M | 4.0 | Too big |
| RTX 3090 24GB | 24 GB | Q4_K_M | 3.8 | Too big |
MacBook Pro M4 Pro 48GB | 48 GB | Q4_K_M | 3.6 | Too big |
| RTX 4080 Super 16GB | 16 GB | Q4_K_M | 3.5 | Too big |
| RTX 4070 12GB | 12 GB | Q4_K_M | 2.2 | Too big |
| RTX 3060 12GB | 12 GB | Q4_K_M | 2.0 | Too big |
| RTX 4060 8GB | 8 GB | Q4_K_M | 2.0 | Too big |
Estimates for single-stream decoding at Q4_K_M; real tokens/sec varies with prompt length, context, batch size, and runtime build. Prompt processing (prefill) is faster than the decode figures shown here.
How Llama 4 Scout 17B 16E (109B params) fits at each quantization level on NVIDIA DGX Spark 128GB (92.2 GB usable).
| Quant | Bits | VRAM | Quality | Fit |
|---|---|---|---|---|
Q2_K | 2 | 42.5 GB | Low | A74 |
Q3_K_S | 3 | 53.4 GB | Low | A76 |
NVFP4 | 4 | 61.0 GB | Medium | A76 |
Q4_K_MBest for your GPU | 4 | 66.5 GB | Medium | A76 |
Q5_K_M | 5 | 78.5 GB | High | F0 |
Q6_K | 6 | 89.4 GB | High | F0 |
Q8_0 | 8 | 116.6 GB | Very High | F0 |
F16 | 16 | 223.5 GB | Maximum | F0 |
Copy-paste commands to run Llama 4 Scout 17B 16E on your machine.
Run
lms load Llama-4-Scout-17B-16E-Instruct && lms server startYour hardware
| Model | Params | Grade | Decode | Capabilities |
|---|---|---|---|---|
 Devstral 2 123B Instruct | 123B | S | 2.4 tok/s | |
 Qwen 3.5 122B A10B | 122B | S | 6.6 tok/s | |
| Mistral Small 4 119B | 119B | S | 7.1 tok/s | |
 GPT-OSS 120B | 117B | A | 2.5 tok/s | |
 Command A 111B | 111B | S | 2.6 tok/s |
Yes, NVIDIA DGX Spark 128GB can run Llama 4 Scout 17B 16E with a A grade (Runs well). Expected decode speed: 6.3 tok/s.
Llama 4 Scout 17B 16E (109B parameters) requires approximately 83.7 GB of memory with Q4_K_M quantization.
The recommended quantization for Llama 4 Scout 17B 16E is Q4_K_M, which balances quality and memory efficiency.
On NVIDIA DGX Spark 128GB, Llama 4 Scout 17B 16E achieves approximately 6.3 tokens per second decode speed with a time-to-first-token of 30909ms using Q4_K_M quantization.
For coding workloads, Llama 4 Scout 17B 16E on NVIDIA DGX Spark 128GB receives a A grade with 6.3 tok/s and 153K context.
On NVIDIA DGX Spark 128GB, Llama 4 Scout 17B 16E can safely use up to 153K tokens of context. The model's official context limit is 10.5M, but available memory constrains the safe maximum.
Prioritize bandwidth, not only capacity. If this workload feels slow, the next useful step is often a GPU tier with materially faster memory bandwidth rather than only a small bump in capacity.
Not always. NVIDIA DGX Spark 128GB can often fit larger models thanks to unified memory, but a discrete GPU with dedicated high-bandwidth VRAM may still decode faster once the model fits. For this combination, the important distinction is capacity versus sustained throughput.
Paste this snippet into any page to show a live fit card.
<iframe src="https://willitrunai.com/embed/llama-4-scout-17b-16e-on-dgx-spark-128gb" width="400" height="180" frameborder="0" style="border:none;border-radius:12px;overflow:hidden;" title="Will It Run AI — fit result"></iframe>
Preview: