LANGUAGE MODEL Meta AI Last updated:

Llama 4

Meta's MoE Open Frontier

Meta's first MoE-native Llama generation, released April 2025 with three sizes — Scout (109B / 17B active, 10M context), Maverick (400B / 17B active), and Behemoth (~2T total, still in preview at launch). Open weights but with stricter licensing than Llama 3 (acceptable use policy expanded). Mixed reception: benchmark scores trailed expectations vs internal Meta commentary that had hyped Behemoth as "GPT-5 class."

Intelligence
Below avg
Speed
Medium
116 tok/s output
Cost
Low
$0.35 in / $0.85 out
Context
10M
Up to 10,000,000 tokens
How are Intelligence, Speed & Cost bucketed?
Intelligence and Speed buckets are percentile ranks on Artificial Analysis. Cost buckets are fixed dollar thresholds keyed off output-token price ($/M out).
Intelligence
  • Top 1%≤ 1%
  • Top 5%≤ 5%
  • Top 10%≤ 10%
  • Good≤ 25%
  • Medium≤ 50%
  • Below avg> 50%
Speed
  • Top 1%≥ 345 tok/s
  • Top 5%≥ 237 tok/s
  • Top 10%≥ 196 tok/s
  • Good≥ 146 tok/s
  • Medium≥ 90 tok/s
  • Slow< 90 tok/s
Cost
  • Freeopen weights · self-host
  • Low< $1 / M out
  • Moderate$1–5 / M out
  • High≥ $5 / M out

Why it matters

Llama 4 illustrates that open-weight leadership is no longer a single-lab story. Meta's structural advantages (compute, data, reach) didn't translate to an obvious win in the generation after Llama 3 — pointing to either a methodological gap or a talent / strategy gap that Meta is now actively trying to close.

Core Capabilities

Long Documents
Handles entire codebases, books, and multi-doc RAG.
Multimodal
Combines text, vision, and audio in one model.
Generative
Produces images, video, audio, or other media.
Agent Workflows
Built for tool use and autonomous tasks.

Context Window

10M tokens
≈ entire codebase
4k Chat 聊天
32k Long docs 长文档
128k Books 整本书
400k Multi-doc 多文档
1M Codebase 整个代码库
10M This model 本模型

Availability

API
Not available
Product / App
Available
Open Source
Released
Enterprise
Contact sales

Pricing Model

Free / self-host
Open weights — pay only for compute.
Self-host

Capability / Performance

Where this model sits relative to the middle 60% of models in the tree. All scores are 0–10 (higher is better).

Lower 20% Upper 80% This model
Reasoning
AA Intelligence Index · scaled to 10
1.7
5.6
2.6
Coding
SciCode · scaled to 10
1.8
4.3
1.7
Agentic tasks
Terminal-Bench Hard · scaled to 10
0.2
3.6
0.1
Context / memory
Context window size · log-scaled
6.0
9.0
10.0
Cost efficiency
Input price ($/M tokens) · cheaper scores higher
6.2
10.0
10.0
Lower 20% 20th percentile — 20% of models score below this This model Where the current model lands Upper 80% 80th percentile — only 20% of models score above this Percentile boundaries are computed across every model in the tree that reports the underlying benchmark for each capability.

What it feels like

  • Meta's first MoE-native Llama and first natively multimodal open-weights generation
  • Maverick (17B active / 128 experts) beats GPT-4o and Gemini 2.0 Flash on most benchmarks at release
  • Scout (17B active / 16 experts) fits in a single H100 with a claimed 10M-token context window
  • Benchmark numbers landed under a cloud — community questioned whether the LMArena score reflected the open-weight checkpoint
  • Real-world testers found gaps between announcement claims and independent reproduction
  • Despite controversy, the open weights gave the ecosystem a viable post-Llama-3 baseline

Best use cases

  • Self-hosted multimodal applications where API models can't go
  • Long-context retrieval and document QA (especially Scout's 10M window)
  • Fine-tuning on private data while staying inside Meta's open license
  • Cost-sensitive multimodal inference at scale

Tools to try

Not ideal for

  • Frontier-leaderboard reasoning — Claude 4, GPT-5, DeepSeek R1 score higher
  • Edge / single-consumer-GPU deployments (even Scout needs an H100)
  • Workflows where the LMArena-style controversy is a credibility risk