How much VRAM to run Llama 4 Scout?

About 76 GB atQ4_K_M with an 8K context — fits a A100 80GB. Full breakdown below, or check your exact hardware.

Llama 4 Scout VRAM by quantisation

QuantisationWeightsTotal (8K ctx)Fits on
Q4_K_M66.5 GB75.9 GBA100 80GB, RTX PRO 6000 Blackwell
Q5_K_M77.6 GB88.1 GBRTX PRO 6000 Blackwell, Apple Silicon 128GB unified
Q6_K89.4 GB101.1 GBNVIDIA DGX Spark (128GB unified), Apple Silicon 256GB unified
Q8_0116.1 GB130.5 GBApple Silicon 256GB unified, Apple Silicon 512GB unified
FP16 / BF16218.0 GB242.6 GBApple Silicon 512GB unified

Check your hardware

About Llama 4 Scout

Llama 4 Scout opened Meta's mixture-of-experts era: a natively multimodal model built with early-fusion vision, a modest active-parameter set drawn from a moderate expert pool, and the generation's headline feature — a context window an order of magnitude beyond anything the industry had shipped, aimed at whole-corpus and multi-document work. Meta's community licence applies as across the Llama line.

Meta positioned Scout explicitly as the single-accelerator member of the pair, the one sized so that serious-but-not-exotic hardware could run a Llama 4 model.

Llama 4 Scout is Meta's 109B-parameter model released in April 2025, with a 10240K-token context window. It is a mixture-of-experts model: all 109B parameters must sit in memory, but only ~17B are active per token, which is what makes it fast for its size. It uses a classic dense-attention design whose KV cache grows linearly with context: its KV cache is about 1.6 GB at an 8K context, 25.8 GB at 128K, and 2061.6 GB at the full 10240K window (FP16 cache).

Scout's memory profile follows MoE rules — the whole expert pool resident, so the totals above track the full count, putting quality quantisations in workstation and big-unified-memory territory rather than gaming-card range. The enormous context window deserves realism: its conventional grouped-query attention means cache grows linearly, so the outer reaches of that window are for high-memory deployments, not a free lunch — set the calculator above to a context you would actually use. Within those bounds it delivers something rare: frontier-generation multimodality on hardware a determined individual can own.

For most people Q4_K_M is the sweet spot — the most popular quality/size trade-off — while Q8 is near-lossless if you have the memory. Totals above include the KV cache and a realistic framework overhead, so they are what you should expect to see in practice rather than just the download size. Weight sizes are calibrated against real GGUF files — see the methodology.

Frequently asked questions

How much VRAM does Llama 4 Scout need?

At Q4_K_M with an 8K context, Llama 4 Scout needs about 76 GB (weights 66 GB + KV cache + overhead). The smallest common hardware that fits is a A100 80GB.

Can an RTX 4090 (24GB) run Llama 4 Scout?

Not fully in VRAM. Llama 4 Scout needs about 76 GB even at Q4_K_M, so a 24 GB card would have to offload layers to system RAM at a large speed penalty.

Can a Mac run Llama 4 Scout?

Yes — Apple Silicon with 128 GB of unified memory or more (macOS lets the GPU use ~75% of it, ~96 GB) runs Llama 4 Scout at Q4_K_M.

How much VRAM does Llama 4 Scout need at its full 10240K context?

The KV cache grows with context: about 1.6 GB at 8K tokens versus 2061.6 GB at the full 10240K window (FP16 cache). Add that to the weights (66 GB at Q4_K_M) plus overhead — long contexts can cost more than a whole quantisation step.

Related

VRAM calculator for any model ·Token counter

Similar-size local models: GPT-OSS 120B (117B) · Mistral Small 4 (119B) · Llama 3.3 70B (70.6B) · DeepSeek V4 Flash (284B) · Qwen3.6 35B-A3B (35B)

Last updated 2026-09-03. Architecture figures from the model's published config.json; see the methodology.