ai.canshu.com
中文
VRAM Calculator Advisor & TCO API Pricing Compare Models Leaderboard GPUs

Search models, GPUs, and tools

↑↓ Navigate · Enter Open · ★ Watchlist prioritized ai.canshu.com
Meta
Open-Weight Meta · MoE 109B / A17B · Text+Image · BF16 / FP8

Llama-4-Scout-109B-A17B (MoE) Architecture Parameters, Benchmarks & VRAM

Meta 109B total / 17B active multimodal MoE tailored for 128GB unified memory workstations.

🧬 Llama-4-Scout-109B-A17B (MoE) Architecture & Tensor Specifications

Verified directly against HuggingFace (meta-llama/Llama-4-Scout-17B-16E-Instruct) config.json & Model Card

HF Class: Llama4ForConditionalGeneration

1. Backbone & Dimensions

llama4
Total / Active Params
109B / 17B
Hidden Size (d_model)
5,120
FFN Intermediate Size
8,192
MoE Experts
16 Routed · Top-1 + 1 Shared
Activation / Norm
SiLU (SwiGLU) · ε=1e-5
Native Tensor Dtype
BF16 / FP8

2. Attention & Layer Layout

GQA
Num Hidden Layers
48 Layers
Hidden Layer Layout
48-Layer iRoPE Interleaved Multimodal MoE (16 Routed + 1 Shared Expert)
Attention Heads (Q / KV)
40 Q / 8 KV (5:1 GQA)
Head Dimension
128
Linear / KV Compression
iRoPE (interleaved RoPE + NoPE temperature scaling) supporting up to 10M context
Vision Encoder / MTP
ViT 34L

3. Context, RoPE & Tokenizer

Tokenizer
Native Context Length
10,485,760 tokens
RoPE Configuration
iRoPE (θ=500,000 · max_position_embeddings=10,485,760 / 10M)
Vocab Size
202,048 (Untied)
Special Token IDs
BOS: 200000 · EOS: [200001,200007,200008] · PAD: 200018

📋 Llama-4-Scout-109B-A17B (MoE) Official API & System Card Specifications

Verified against official Meta developer documentation & pricing

API Model ID: meta-llama/llama-4-scout

1. Identity & Capabilities

API Spec
API Model ID
meta-llama/llama-4-scout
Release Date
2025-04-05
Weight License
Open-Weight (API Spec)
Native Reasoning
—
Function Calling & JSON Schema
✅ 原生支持
Batch API 异步批处理
✅ Supported (50% off)

2. Context, Output & Modalities

Limits
Max Context Window
1,024,000 tokens (1024K)
Max Output Tokens
16,000 tokens
Input Modalities
Text + Image
Output Modalities
Text
Direct China Access
✅ Direct CN + RMB
Deployment Mode
Open-Weight + Cloud API

3. Official Pricing & Cache

USD / 1M
Input / 1M Tokens
$ 0.1
Cached Input / 1M
$ 0.1
Output / 1M Tokens
$ 0.3
Tiered Pricing Notes
OpenRouter: Unified multi-provider routing via OpenRouter

🏆 Official Benchmark Results

Sourced from official System Card / Model Card & technical reports

Arena Elo ~1438 AA Index 42
📚 Knowledge
MMLU 79.6 %

Massive Multitask Language Understanding benchmark testing knowledge across 57 diverse subjects including STEM, humanities, social sciences, and professional domains.

📚 Knowledge
MMLU-Pro 74.3 %

A more robust and challenging multi-task language understanding benchmark that extends MMLU by expanding multiple-choice options from 4 to 10, eliminating trivial questions, and focusing on reasoning-intensive tasks.

🧠 Reasoning
GPQA Diamond 57.2 %

PhD-level Google-proof science reasoning questions

🧠 Reasoning
MATH 50.3 %

MATH dataset contains 12,500 challenging competition mathematics problems from AMC 10, AMC 12, AIME, and other mathematics competitions.

👁️ Multimodal
MathVista 70.7 %

MathVista evaluates mathematical reasoning of foundation models in visual contexts. It consists of 6,141 examples derived from 28 existing multimodal datasets and 3 newly created datasets (IQTest, FunctionQA, and PaperQA), combining challenges from diverse mathematical and…

👁️ Multimodal
MMMU 69.4 %

MMMU (Massive Multi-discipline Multimodal Understanding) is a benchmark designed to evaluate multimodal models on college-level subject knowledge and deliberate reasoning.

💻 Coding
LiveCodeBench 32.8 %

LiveCodeBench is a holistic and contamination-free evaluation benchmark for large language models for code.

🎛️ Official Sampling & Reasoning Parameters

Recommended generation settings & API parameter constraints

Official Default Multimodal Chat
temperature=0.6 top_p=0.9 max_tokens=16,384

📊 Llama-4-Scout-109B-A17B (MoE) Full Benchmark Scorecard

11 public benchmarks · ranks among 154 models · synced 2026-11-20

⚔️ Compare head-to-head

Source: ZeroEval / llm-stats.com (✓ = third-party verified, others vendor-reported).

⚡ Llama-4-Scout-109B-A17B (MoE) Interactive VRAM & Token Speed Calculator

Concurrency (Batch):
MoE 109B / A17B·48L (d=5120)·40Q/8KV·Ctx 1M·Text+Image
Runnable Models →
128 GB Unified·614 GB/s·Safe Usable 108.8 GB
Smooth Fit71.92 / 108.8 GB~34 tok/s
3. Weight Quantization
4. Per-Request Context16K (16,384 tok)
Verified: 2026-11-20·Roofline bandwidth/compute estimate (Batch=1)
Smooth Fit
71.92/ 108.8 GB usable

Physical 128 GB · Load 66.1%

Theoretical Decode Speed
~34tokens/s

MoE: reads only 10.31 GB active weights/step

Max Context at Batch=1
199K/ 1M

Enable Q8/Q4 KV quantization to scale batch concurrency

Weights: 66.08 GBKV Cache (1×16K): 3.22 GBOverhead: 2.62 GB
Reserves ~15% unified memory for OS
🏃 Smooth & Snappy~14.7s for a 500-token response
Reference Launch CLI:
OLLAMA_CONTEXT_LENGTH=16384 ollama run llama4:scout

Mac Studio M5 Max 128G Fits Llama-4-Scout-109B-A17B (MoE) · Cost & TCO

Approx. $3,499 (~$27/GB VRAM) · 120W TDP

📊 Llama-4-Scout-109B-A17B (MoE) Quantization VRAM Breakdown (16K Context)

Quantization Bits/Weight Weights KV Cache (16K) Total VRAM Quality Note
FP16 / BF16 16 bpw 218 GB 3.22 GB 226.88 GB Lossless Full Precision
FP8 / Q8_0 8.5 bpw 115.81 GB 3.22 GB 122.65 GB 99.9% Near-Lossless
GGUF Q6_K 6.56 bpw 89.38 GB 3.22 GB 95.69 GB Ultra High Quality
GGUF Q5_K_M 5.69 bpw 77.53 GB 3.22 GB 83.6 GB High Balanced Quality
Q4_K_M / AWQ 4.85 bpw 66.08 GB 3.22 GB 71.92 GB Sweet Spot (Recommended)
GGUF IQ4_XS 4.25 bpw 57.91 GB 3.22 GB 63.59 GB Compact 4-bit
GGUF Q3_K_M 3.91 bpw 53.27 GB 3.22 GB 58.86 GB Moderate Perplexity Loss
GGUF Q2_K 3.15 bpw 42.92 GB 3.22 GB 48.3 GB Extreme Compression

🖥️ GPU & Apple Silicon Compatibility for Llama-4-Scout-109B-A17B (MoE) (Q4_K_M)

Frequently Asked Questions

What are the exact architecture parameters of Llama-4-Scout-109B-A17B (MoE)?

Llama-4-Scout-109B-A17B (MoE) uses Llama4ForConditionalGeneration with 48 layers (48-Layer iRoPE Interleaved Multimodal MoE (16 Routed + 1 Shared Expert)), hidden_size=5120, 40 Q heads / 8 KV heads (head_dim=128), and vocab_size=202,048.

How much VRAM does Llama-4-Scout-109B-A17B (MoE) require?

At 4-bit quantization (Q4_K_M, 16K context), Llama-4-Scout-109B-A17B (MoE) requires 71.92 GB VRAM. Full FP16 requires 226.88 GB.