The Open-Source AI Stack
RSS
All models

Models · llama

Llama 4 Maverick

Source-available Meta · 2025-04-05 · Llama 4 Community License

The mid-tier sibling in Meta's first MoE Llama family, sharing the natively multimodal early-fusion architecture used across Llama 4. Maverick pairs 17B active parameters with 128 routed experts, and Meta positioned it against GPT-4o and Gemini 2.0 Flash on standard reasoning and coding suites.

Cost

$0.35 / Mtok input
$0.85 / Mtok output

Together AI · as of 2026-05-21

via Artificial Analysis ↗

Speed

108.9 tok/sec output
682 ms TTFT

· as of 2026-05-21

source ↗

Architecture

tokens in Embedding vocab not disclosed · llama4 tokenizer × N layers Grouped-Query Attention RoPE (Llama 3 scaling) context 1,048,576 tokens MoE Router 128 experts total · 1 active per token shown: 32 of 128 Output projection tokens out
Schema-generated from data/models.yaml. Every label is auditable against the model's sources.

Specs

Architecture
moe
Total params
400B
Active params
17B
Experts
128 total · 1 active
Context window
not verified
Attention
gqa
Position encoding
rope-llama3
Pretraining tokens
22.0T
Training hardware
H100
Post-training
sft, online-rl, dpo
OSI-approved
no
Data released
no
Training code
not released

Benchmarks

Each score carries the date it was published; we never infer or interpolate missing scores.

Math

MATH 88.9 as of 2026-05-21 source ↗
AIME 2024 39.0 as of 2026-05-21 source ↗
AIME 2025 19.3 as of 2026-05-21 source ↗

Held-out / arena

LMArena Elo 1417.0 as of 2025-04-05 source ↗

Available quantizations

GGUF llama.cpp's container; the common local format, k-quants from Q2 to Q8. runs on llama.cpp, Ollama
AWQ Activation-aware 4-bit weight quantization for GPU serving. runs on vLLM, SGLang
MLX Apple MLX 4/8-bit layout for Apple silicon. runs on Apple MLX
FP8 8-bit float, frequently a native release on Hopper / Blackwell GPUs. runs on vLLM, SGLang, TensorRT-LLM

Verified via the Hugging Face model tree ↗. Community quantizations change over time; the families shown are those with published weights at audit time.

Notable innovations

  • · 128 routed experts plus shared expert
  • · Native multimodal early fusion
  • · Online RL post-training loop

Sources