LANGUAGE MODEL Meta AI Last updated:

Llama 3

Meta's Frontier-Comparable Open Model

Meta's third-generation open-weight model, trained on 15 trillion tokens — over 7× more than Llama 2. The 70B variant matched or approached GPT-4 and Claude 3 Sonnet on most benchmarks while being downloadable for free with a permissive license. Followed in July 2024 by Llama 3.1 with a 405B version that explicitly targeted GPT-4 / Claude 3.5 parity.

Intelligence
Below avg
Speed
Slow
47 tok/s output
Cost
Moderate
$0.65 in / $2.75 out
Context
8K
Up to 8,192 tokens
How are Intelligence, Speed & Cost bucketed?
Intelligence and Speed buckets are percentile ranks on Artificial Analysis. Cost buckets are fixed dollar thresholds keyed off output-token price ($/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 3's release cadence and quality made open weights the structural baseline for LLM economics. Any closed model now has to justify its premium against a free open alternative within one to two quarters of release. This pricing pressure is the single biggest factor reshaping the LLM business in 2024-26.

Core Capabilities

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

Context Window

8k tokens
≈ short doc
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
8k

Availability

API
Not available
Product / App
Not available
Open Source
Released
Enterprise

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
1.3
Coding
SciCode · scaled to 10
1.8
4.3
2.6
Agentic tasks
Terminal-Bench Hard · scaled to 10
0.2
3.6
0.3
Context / memory
Context window size · log-scaled
6.0
9.0
2.0
Cost efficiency
Input price ($/M tokens) · cheaper scores higher
6.2
10.0
9.3
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

  • Major leap over Llama 2 — 70B operates in the same 'weight class' as GPT-4 and Claude 3 Sonnet at release
  • Climbed to 5th place on the LMSYS leaderboard within weeks — first open-weights model to genuinely compete with GPT-4
  • 128K-token vocabulary tokenizer noticeably improved encoding efficiency vs Llama 2
  • Trained on 15T tokens (7x Llama 2) with 4x more code — coding feel is dramatically better
  • Strong human-preference rankings vs comparable closed models in real-world scenarios
  • 8B variant sets a new bar for laptop-class quality, especially after instruction tuning

Best use cases

  • Open-weights production deployments where Llama 2 was the previous baseline
  • Self-hosted RAG, code-completion, and chat applications under Llama 3 license
  • Fine-tuning on private corpora (the open license makes vertical specialisation safe)
  • Research and academic work that needs an open frontier-tier baseline

Tools to try

Not ideal for

  • Frontier reasoning leaderboards by late 2025 — superseded by Llama 3.1/3.3/4 and competitors
  • Multimodal tasks — text-only at this generation (Llama 3.2 Vision came later)
  • Workflows blocked by Meta's acceptable-use clauses (>700M MAU restriction, etc.)

Model Evolution

View full evolution tree →