llmrun
MetaLlama 4

Llama 4 Scout 17B 16E Instruct — Hardware Requirements & GPU Compatibility

Vision

Llama 4 Scout is the smaller of Meta's first Llama 4 models, a natively multimodal Mixture-of-Experts model with 16 experts, about 109 billion total parameters and 17 billion active per token. This instruction-tuned checkpoint accepts text and images and generates text, and is aimed at assistant-style chat, visual understanding and long-document work. Meta advertises an unusually long context window of up to 10 million tokens, and the model is released under the Llama 4 Community License. Because every expert has to stay in memory, 4-bit quantization still needs roughly 60GB, which means a high-memory workstation, a large unified-memory Mac or a multi-GPU setup rather than a single consumer card.

108.6B params · MoE17B activeOther162.7K downloads1.4K likes40.9K quant downloads

Llama 4 Scout 17B 16E Instruct runs on a 96 GB+ Mac or a workstation GPU

Best local pick → Mac Studio (M3 Ultra, 96GB) · Q4_K_M · 77 GB · ~34 tok/s

Macs from48 GB IQ2_S · ~43 tok/s

Score 81.8#79 of 100 open#40 of 60 that fit ≤96 GBOther

Not set up for it?

Reference GPU · RTX 4090 24 GB

Estimates, ±20% · how we estimate

llmrun Composite Score →

81.8

Benchmarks · tier chart ↓
Among 100 open modelsHow the Composite Score is built

Run Llama 4 Scout 17B 16E Instruct

One recommended way. Every other method, OS and quant is in the step-by-step guide.

Other ways

each opens step-by-step

Run in cloud

Fits on RTX PRO 6000 (96 GB) (19 GB headroom) · Q4_K_M

Generation speed
~74 tok/s
generation speed
Cost per 1M output tokens
$5.73
per 1M output tokens
Deployon Vast.aiBest optionfrom $1.54/hr (opens in new tab)
Compare GPUs →
or

How Much VRAM Does Llama 4 Scout 17B 16E Instruct Need?

Select a quantization level; the listed hardware adheres to it.

QuantizationVRAM+32KOn RTX 4090 (reference)
50.8 GB62.2 GB2-bit · K-quant — 2-bit quantization with K-quant improvementsDoes not fit: needs 56.1 GB
52.3 GB63.7 GB3-bit small — 3-bit small quantizationDoes not fit: needs 57.6 GB
58.3 GB69.7 GB3-bit medium — 3-bit medium quantizationDoes not fit: needs 63.6 GB
59.8 GB71.2 GB4-bit · legacy — 4-bit legacy quantizationDoes not fit: needs 65.1 GB
71.7 GB83.1 GB4-bit medium — 4-bit medium quantization — most popular sweet spotDoes not fit: needs 77 GB
85.2 GB96.6 GB5-bit medium — 5-bit medium quantization — good quality/size tradeoffDoes not fit: needs 90.5 GB
98.6 GB110 GB6-bit, very good — 6-bit quantization, very good qualityDoes not fit: needs 103.9 GB
119.5 GB130.9 GB8-bit, near-lossless — 8-bit quantization, near-losslessDoes not fit: needs 124.8 GB

Selected Q4_K_M

4-bit medium quantization

77.0GBat 16K context

Weights
70.6 GB
KV cache
6.1 GB
Overhead
0.3 GB

Doesn't fit RTX 4090 (reference) · needs 53.0 GB more

≥30 tok/s needs ≥561 GB/s · M3 Ultra ~34, M2 Ultra ~34

Or rent RTX PRO 6000 from $1.54/hr.

Which Hardware Can Run Llama 4 Scout 17B 16E Instruct?

Q4_K_M · 71.7 GB

Llama 4 Scout 17B 16E Instruct (Q4_K_M) requires 71.7 GB of VRAM to load the model weights. For comfortable inference with headroom for KV cache and system overhead, 94+ GB is recommended. No consumer GPU has enough memory. 30 devices with unified memory can run Llama 4 Scout 17B 16E Instruct, including Mac Studio (M3 Ultra, 256GB), Mac Studio (M3 Ultra, 512GB), Mac Studio (M3 Ultra, 96GB).

By VRAM tier · typical GPUs and Macs at Q4_K_M
8–16 GB · RTX 4060, RTX 3060 12GB, RTX 5060 Ti 16GBDoesn't fit — smallest quant needs 38.2 GB8 GB tier →12 GB tier →16 GB tier →
24 GB · RTX 4090referenceDoesn't fit — smallest quant needs 38.2 GB24 GB tier →
Mac 32 GB · Mac Mini M4Doesn't fit — smallest quant needs 38.2 GBMacs →
Mac 64 GB · Mac Studio M4 MaxIQ3_M · 59.1 GB · ~32 tok/s · 16KMacs →

How good is Llama 4 Scout 17B 16E Instruct?

The llmrun Score merges 11 public benchmark boards into a single figure, where 100 represents the average for open models.

Score vs VRAM in the ≤96 GB tier

  • Llama 4 Scout 17B 16E Instruct
  • Other models
  • Best Score for its VRAM

#40 of 60 that fit ≤96 GB

Each dot is a model at the best quant that fits ≤96 GB. The line joins the models with the best Score for their VRAM; a dot below it is beaten by a model that needs no more memory.

Full ≤96 GB ranking →

Benchmarks

The 6 boards that weigh most in the llmrun Score, of 13 with a result for this model; badges show who ran the test.

  • GPQA Diamond

    Epoch AI · Apr 2025Independent

    51.8%

    #51/98 open

    Independent
  • LMArena Text

    LMArena · Oct 2026Arena

    1321.5

    #94/187 open

    Arena
  • AIME 2024/2025

    Epoch AI · Apr 2025Independent

    7.8%

    #62/86 open

    Independent
  • ARC-AGI-2

    Epoch AI · Apr 2025Independent

    0.0%

    #18/19 open

    Independent
  • MMLU-Pro

    TIGER-LabIndependent

    74.3%

    #31/145 open

    Independent
  • SciCode

    Epoch AI · Apr 2025Lab-reported

    17.0%

    #63/66 open

    Lab-reported
All 13 Llama 4 Scout 17B 16E Instruct benchmarks

In the ≤96 GB class

60 models fit ≤96 GB

Same Score, only models that fit ≤96 GB at their best quant: the ranking a ≤96 GB owner actually chooses from.

  1. #1

    MiMo V2.6 Flash RLXiaomi

    131.9

    ~153 tok/s · 93.8 GB

  2. #2

    DeepSeek V4 Flash 0731DeepSeek

    127.6

    ~103 tok/s · 95.5 GB

  3. #3

    Qwen3.8 27BAlibaba

    122.1

    ~67 tok/s · 17.9 GB

  4. #4

    Inkling SmallThinking Machines

    119.6

    ~356 tok/s · 92.9 GB

  5. #5

    MiniMax M2.5MiniMax

    111.8

    ~138 tok/s · 93.9 GB

  6. #6

    Qwen3.6 27BAlibaba

    111.2

    ~67 tok/s · 17.9 GB

Full ≤96 GB ranking →RTX PRO 6000 Blackwell Workstation Edition page →

Alternatives to Llama 4 Scout 17B 16E Instruct

Official models with an llmrun Score, each answering one question; numbers are at the best quant that fits a 24 GB GPU.

Similar models

10
ModelScoreVRAMtok/sWhy it is similar
MiniMax M2.5MiniMax111.8139.6 GB—needs > 24 GBsimilar score
Qwen3.5 122B A10BAlibaba109.475.5 GB—needs > 24 GBsimilar score
MiniMax M2.7MiniMax107.4139.6 GB—needs > 24 GBsimilar score
MiniMax M2.1MiniMax106.2139.6 GB—needs > 24 GBsimilar score
MiniMax M2MiniMax104.0139.6 GB—needs > 24 GBsimilar score
Qwen3 235B A22BAlibaba99.6142.9 GB—needs > 24 GBsimilar score
GPT OSS 20BOpenAI96.313.1 GB~190similar score
Phi 4Microsoft83.912.5 GB~69similar score
Qwen2.5 72B InstructAlibaba81.749.3 GB—needs > 24 GBsimilar score
Llama 3.3 70B InstructMeta79.751.3 GB—needs > 24 GBsimilar score
Fine-tunes & variants(1)

Specifications

Publisher
Meta
Family
Llama 4
Parameters
108.6B
Active parameters
17B per token
Release Date
2025-04-02
License
Other

Resources

Frequently Asked Questions

How much VRAM does Llama 4 Scout 17B 16E Instruct need?

Llama 4 Scout 17B 16E Instruct requires 71.7 GB of VRAM at Q4_K_M, or 239.0 GB at BF16.

VRAM = Weights + KV Cache + Overhead

Weights = 108.6B × 4.8 bits ÷ 8 = 65.2 GB

KV Cache + Overhead ≈ 6.5 GB (at 2K context + ~0.3 GB framework)

Fit ratings and hardware model lists check this model with room for a 16K-token context, which needs a little more memory.

VRAM usage by quantization

71.7 GB

Learn more about VRAM estimation →

Can NVIDIA GeForce RTX 5090 run Llama 4 Scout 17B 16E Instruct?

No — Llama 4 Scout 17B 16E Instruct requires at least 32.9 GB at IQ2_XXS, which exceeds the NVIDIA GeForce RTX 5090's 32 GB of VRAM.

What's the best quantization for Llama 4 Scout 17B 16E Instruct?

For Llama 4 Scout 17B 16E Instruct, Q4_K_M (71.7 GB) offers the best balance of quality and VRAM usage. Q5_K_S (82.2 GB) provides better quality if you have the VRAM. The smallest option is IQ2_XXS at 32.9 GB.

VRAM requirement by quantization

IQ2_XXS
32.9 GB
IQ3_XS
49.3 GB
Q4_0
59.8 GB
IQ4_NL
67.2 GB
Q4_K_M ★
71.7 GB
BF16
239.0 GB

★ Recommended — best balance of quality and VRAM usage.

Learn more about quantization →

Can I run Llama 4 Scout 17B 16E Instruct on a Mac?

Yes — Mac Studio M5 Max (48 GB) and 55 other Macs can run Llama 4 Scout 17B 16E Instruct, the smallest at IQ2_S. Apple Silicon uses unified memory, so the model shares RAM with the system. At Q4_K_M you need at least 71.7 GB of usable unified memory (RAM minus macOS overhead), from Mac Studio (M3 Ultra, 96GB).

Can I run Llama 4 Scout 17B 16E Instruct locally?

Yes — Llama 4 Scout 17B 16E Instruct can run locally on consumer hardware. Popular tools include Ollama, LM Studio, and llama.cpp.

How fast is Llama 4 Scout 17B 16E Instruct?

At Q4_K_M, Llama 4 Scout 17B 16E Instruct can reach ~84 tok/s on AMD Instinct MI350X. Speed depends mainly on GPU memory bandwidth. Real-world results typically within ±20%.

tok/s = 1000 ÷ (active GB ÷ (bandwidth GB/s × efficiency) × 1000 + layers × routing ms)

Mixture-of-Experts: just the active experts get read for each token, along with a fixed routing cost per layer.

Example: NVIDIA B200 → 11.2 GB active ÷ (8000 × 0.65) = 2.16 ms, plus 48 layers × 0.055 ms = 2.64 ms, so 1000 ÷ 4.80 ms = ~208 tok/s

Estimated speed at Q4_K_M (71.7 GB)

~208 tok/s
~208 tok/s
~162 tok/s

Real-world results typically within ±20%. Speed depends on batch size, quantization kernel, and software stack.

Learn more about tok/s estimation →

What's the download size of Llama 4 Scout 17B 16E Instruct?

At Q4_K_M, the download is about 65.19 GB. The full-precision BF16 version is 217.28 GB. The smallest option (IQ2_XXS) is 29.88 GB.

Which GPUs can run Llama 4 Scout 17B 16E Instruct?

No single consumer GPU has enough VRAM to run Llama 4 Scout 17B 16E Instruct at Q4_K_M (71.7 GB). Multi-GPU or professional hardware is required.

Which devices can run Llama 4 Scout 17B 16E Instruct?

30 devices with unified memory can run Llama 4 Scout 17B 16E Instruct at Q4_K_M (71.7 GB), including ASUS Ascent GX10, Beelink GTR9 Pro (Ryzen AI Max+ 395, 128 GB), Framework Desktop (Ryzen AI Max+ 395, 128 GB), GMKtec EVO-X2 (Ryzen AI Max+ 395, 128 GB). Apple Silicon Macs use unified memory shared between CPU and GPU, making them well-suited for local LLM inference.