How much VRAM Llama 4 Scout 17B-16E (MoE) needs, and what it costs to run
Some links on this page are referral links: we may earn a commission, at no cost to you.
Trying to decide what to buy? The best GPU for Llama 4 Scout 17B-16E (MoE) →
Cheapest way to rent it right now
$0.13/hr
NVIDIA DGX Spark (GB10 Grace Blackwell) · Q8_0 · spot · ≈ $32/mo at 8h/day
Live price captured Sat, 14 Nov 2026 01:20:08 GMT.It never changes which card is cheapest.
Just want to use it? Per-token API
$0.18in·$0.59out
per 1M tokens · Llama 4 Scout Instruct on Novita AI · read Oct 29, 2026
No GPU to manage: you pay per token, which suits occasional use. Your prompts run on Novita's servers, not your own machine.
Price as listed on Novita's pricing page on Oct 29, 2026; it may have changed.
Rent a GPU or pay per token?
A per-token API charges for every token, so it is cheapest when you use Llama 4 Scout 17B-16E (MoE) a little. A rented GPU charges by the hour whether it is busy or idle, so it only pays off once you push enough tokens through it. Set your daily volume and see where the line falls for this model, using the cheapest live rental that fits and that we have a memory-bandwidth figure for (DGX Spark (GB10 Grace Blackwell), Q8_0, $0.133/hr spot) and Novita AI's published rate.
- Pay per token · Novita AI$8.47/mo
$0.28/day · $0.28 per 1M tokens blended at 3:1
- Rent DGX Spark (GB10 Grace Blackwell) · Vast.ai$31.85/mo
$1.06/day · 8.0 GPU-hours needed
At 1M tokens/day the API is cheaper. The API is cheaper at every volume: at our estimated speed, renting the DGX Spark (GB10 Grace Blackwell) works out to about $1.06 per 1M tokens against the API's $0.28.
At a glance · 3:1 input:output
| Tokens/day | API /mo | Rent /mo |
|---|---|---|
| 100K | $0.85 | $4.00 |
| 1M | $8.47 | $31.85 |
| 10M | $84.75 | $318 |
| 100M | $848 | $3,185 |
Cheaper option in green. 30-day month. Break-even: none, the API stays cheaper.
How we work it out. API cost = input tokens × $0.18 + output tokens × $0.59 per 1M. Rent cost = the hourly price × the GPU time needed, with at least 1 hour billed per day (spin-up, model load and idle time cost money too). GPU time assumes the DGX Spark (GB10 Grace Blackwell) writes ~11 output tokens/sec for Llama 4 Scout 17B-16E (MoE) at Q8_0 (our estimate: its 273 GB/s memory bandwidth from the spec sheet, not yet confirmed by a second source ÷ the bytes of weights read per token, × 75% efficiency, one request at a time) and reads prompt tokens 10× faster than it writes. Above 24 GPU-hours a day you would need more than one card.
These are estimates: the tokens/sec figure is our model, not a benchmark, and batching many requests (vLLM, SGLang) can raise a rented card's throughput a lot, which moves the break-even lower. Rental price captured Sat, 14 Nov 2026 01:20:08 GMT; it is a spot price, and spot machines can be interrupted. API price as listed by Novita AI on Oct 29, 2026; API prices change.
Memory needed, by quantisation
At 16k context. Weights + KV cache + ~0.8 GB overhead. Memory is in binary GB (1,024³ bytes), the unit GPU memory is sold in; Hugging Face and Ollama list downloads in decimal GB, about 7% more.
| Quant | Weights | KV cache | Total | Quality |
|---|---|---|---|---|
| F16 | 202.4 GB | 0.8 GB | 203.9 GB | Full precision. Reference quality, twice the size of Q8 for no practical gain in most chat use. |
| Q8_0 | 107.5 GB | 0.8 GB | 109.0 GB | Effectively lossless. Use when VRAM is not the constraint. |
| Q6_K | 83.0 GB | 0.8 GB | 84.5 GB | Very close to Q8 at meaningfully less memory. A safe high-quality pick. |
| Q5_K_M | 71.8 GB | 0.8 GB | 73.4 GB | Small, hard-to-notice quality loss. Good balance. |
| Q5_0 | 70.1 GB | 0.8 GB | 71.6 GB | Older-style 5-bit. Q5_K_M is usually the better pick at the same size. |
| Q4_K_M | 61.3 GB | 0.8 GB | 62.9 GB | The community default. Best quality-per-gigabyte for most people. |
| Q4_0 | 57.5 GB | 0.8 GB | 59.1 GB | Older-style 4-bit, measurably worse than Q4_K_M at a similar size. Avoid unless required. |
| IQ4_XS | 53.8 GB | 0.8 GB | 55.3 GB | Newer 4-bit, smaller than Q4_K_M with comparable quality. Needs a recent llama.cpp. |
| Q3_K_M | 49.5 GB | 0.8 GB | 51.0 GB | Noticeable degradation. Use to fit a larger model that would otherwise not run. |
| Q2_K | 42.4 GB | 0.8 GB | 43.9 GB | Heavy degradation. Almost always better to run a smaller model at Q4_K_M instead. |
| IQ3_XS | 41.7 GB | 0.8 GB | 43.3 GB | Aggressive. Usually better than a smaller model at Q4, but test before trusting it. |
Cards that can run it
Best quantisation each card fits at 16k context, with live rental price where we track one.
| GPU | VRAM | Best fit | Uses | ~tok/s | Rent from |
|---|---|---|---|---|---|
| Apple M5 Ultra | 512 GB | F16 | 204 GB | — | — |
| Apple M3 Ultra | 512 GB | F16 | 204 GB | — | — |
| Apple M2 Ultra | 192 GB | Q8_0 | 109 GB | — | — |
| AMD Ryzen AI Max+ 395 (Strix Halo, 128 GB) | 128 GB | Q8_0 | 109 GB | — | — |
| Apple M4 Max (16-core CPU / 40-core GPU) | 128 GB | Q8_0 | 109 GB | — | — |
| NVIDIA DGX Spark (GB10 Grace Blackwell) | 128 GB | Q8_0 | 109 GB | — | $0.133/hrVast.ai |
| Apple M5 Max (18-core CPU / 40-core GPU) | 128 GB | Q8_0 | 109 GB | — | — |
| NVIDIA RTX PRO 6000 Blackwell Workstation Edition | 96 GB | Q6_K | 84 GB | — | $0.400/hrVast.ai |
| NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition | 96 GB | Q6_K | 84 GB | — | $0.271/hrVast.ai |
| Apple M2 Max | 96 GB | Q6_K | 84 GB | — | — |
| Apple M4 Pro | 64 GB | Q4_K_Mtight | 63 GB | — | — |
| AMD Radeon PRO W7900 | 48 GB | Q2_Ktight | 44 GB | — | — |
| NVIDIA RTX 6000 Ada Generation | 48 GB | Q2_Ktight | 44 GB | — | $0.250/hrVast.ai |
| NVIDIA RTX A6000 | 48 GB | Q2_Ktight | 44 GB | — | $0.133/hrVast.ai |
How these numbers are produced
- Model shape comes from the model's own config.json — 12 layers, 40 attention heads, 8 KV heads.
- Weight size is parameters × effective bits-per-weight. Those constants are checked against real published quantised file sizes — median error 0.7% across 40 measurements.
- KV cache is 2 × layers × KV-heads × head-dim × context × 2 bytes. Using KV-heads rather than attention heads is what makes this correct for grouped-query attention; treating a GQA model as multi-head overstates the cache by up to 8×.
- Tokens/sec is an estimate, not a benchmark. Generation is memory-bandwidth-bound, so this is bandwidth ÷ weight-bytes derated to 75%. Real throughput depends on your runtime, batch size and kernels. We don't run our own hardware tests — see our editorial policy.
- This is a mixture-of-experts model. Generation reads only the experts active for each token (1 of 16), so throughput is scored on ~17.0B active parameters, not the full 109B. Memory is the opposite: every expert must still be resident in VRAM, so the figures above use the full weight set.Hybrid MoE: 48 layers, 3 of 4 use chunked attention (8,192-token chunks: 36 layers × 8 KV heads × 128, fixed ~1.2 GB at fp16) and every 4th is a global NoPE layer; `layers`/kv_heads/head_dim = the 12 global layers. Every layer is MoE: 16 routed experts, 1 routed + 1 shared per token; active = vendor-published 17B, but all ~109B weights (vision included) stay resident. Max context 10M per config. Official repo is gated: config.json read 4 Oct 2026 from the unsloth/Llama-4-Scout-17B-16E-Instruct mirror (safetensors total identical to the official API figure).
- Some memory-bandwidth figures are not yet independently verified.Where that's the case we show no tokens/sec at all rather than a number we can't stand behind.
- Rental prices are pulled hourly from provider APIs (last updated Sat, 14 Nov 2026 01:20:08 GMT). Spot/interruptible pricing can change or vanish without notice.
Try it interactively
Drag the context slider and watch the KV cache fill the card
Worth reading before you buy
the cheapest way to run Llama 4 in the cloud: Scout on one H100, Maverick only if you must
Scout fits a single H100 from ~$1.73/hr. Maverick needs 243GB — four H100s or two H200s. Here's the cheapest cloud setup for each, dated to real prices.
Llama 4 vs Qwen3 for Local Use: The Honest Comparison
The quiet truth: for most home users, Qwen3 is the better default than Llama 4 Scout. Where each actually wins, the LMArena controversy, and the specific cases where Scout is worth it.
How much VRAM do you need?
About 0.57 GB per billion parameters at 4-bit, plus room for context: the rule behind the numbers on this page, with worked examples.
The best GPU for local AI right now
The card with enough VRAM for your model at the lowest price, budget by budget.