How much VRAM to run Llama 4 Maverick?

About 271 GB atQ4_K_M with an 8K context — fits a Apple Silicon 512GB unified. Full breakdown below, or check your exact hardware.

Llama 4 Maverick VRAM by quantisation

QuantisationWeightsTotal (8K ctx)Fits on
Q4_K_M244.0 GB271.2 GBApple Silicon 512GB unified
Q5_K_M284.8 GB316.1 GBApple Silicon 512GB unified
Q6_K328.0 GB363.6 GBApple Silicon 512GB unified
Q8_0426.0 GB471.4 GBmulti-GPU / datacenter only
FP16 / BF16800.0 GB882.8 GBmulti-GPU / datacenter only

Check your hardware

About Llama 4 Maverick

Llama 4 Maverick is the larger workhorse of Meta's first mixture-of-experts generation: the same active-parameter count as its Scout sibling, drawn from a much larger expert pool — many smaller experts rather than few big ones — with native early-fusion multimodality and a very long context window. Meta framed it as the general-purpose flagship of the pair, aimed at assistant and creative workloads, under the Llama community licence.

The design bet is legible in the shape: keep per-token compute constant, scale quality through expert diversity, and let deployment cost land where it lands.

Llama 4 Maverick is Meta's 400B-parameter model released in April 2025, with a 1024K-token context window. It is a mixture-of-experts model: all 400B 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 206.2 GB at the full 1024K window (FP16 cache).

That large expert pool is the whole planning problem: identical speed characteristics to Scout, several times the resident memory, which the totals above translate into multi-GPU servers and only the largest unified-memory machines at aggressive quantisation. Nothing consumer-class applies. For most readers this page is a build-versus-API comparison — the same weights are served by many hosted providers, and self-hosting pencils out only at sustained volume on owned or rented accelerators. If Maverick's quality tier is the goal on lesser hardware, the practical route is Scout plus patience, not Maverick plus offloading.

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 Maverick need?

At Q4_K_M with an 8K context, Llama 4 Maverick needs about 271 GB (weights 244 GB + KV cache + overhead). The smallest common hardware that fits is a Apple Silicon 512GB unified.

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

Not fully in VRAM. Llama 4 Maverick needs about 271 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 Maverick?

Yes — Apple Silicon with 512 GB of unified memory or more (macOS lets the GPU use ~75% of it, ~384 GB) runs Llama 4 Maverick at Q4_K_M.

How much VRAM does Llama 4 Maverick need at its full 1024K context?

The KV cache grows with context: about 1.6 GB at 8K tokens versus 206.2 GB at the full 1024K window (FP16 cache). Add that to the weights (244 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: Qwen3.5 397B-A17B (397B) · DeepSeek V4 Flash (284B) · Mistral Large 3 (675B) · GLM-5.2 (753B) · Mistral Small 4 (119B)

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