Question 1

What are all the Llama models from Meta?

Accepted Answer

Meta’s open-weights Llama family tracked here covers three sizes: Llama 3.1 8B (small, fast), Llama 3.3 70B (the production sweet spot) and Llama 3.1 405B (the largest open-weights frontier model). All are released under the Llama Community License — free for nearly all commercial use.

Question 2

What is the latest Llama model?

Accepted Answer

Llama 3.3 70B (December 2024) is the most recent release in the lineup we track. Meta typically refreshes the 70B tier first; a Llama 4 family is widely expected through 2026.

Question 3

Which is the best Llama model?

Accepted Answer

Llama 3.3 70B is the best practical choice — it matches Llama 3.1 405B on most benchmarks at a fraction of the inference cost (intelligence index 66 vs 64). For extreme quality demands you can still use the 405B; for very high-volume traffic the 8B model is the cheap default.

Question 4

Is Llama really open source?

Accepted Answer

Yes — the weights are publicly downloadable and free to use for nearly all commercial cases. There is one caveat: the Llama Community License restricts companies with more than 700 million monthly active users (effectively Meta’s largest competitors). For everyone else it behaves like an open-source license.

Question 5

Can I run Llama on my own hardware?

Accepted Answer

Yes. Llama 3.1 8B runs on a single consumer GPU (RTX 3090 / 4090 / Mac M-series). Llama 3.3 70B needs 2–4× A100 80GB or equivalent. Llama 3.1 405B needs an 8×H100 node. Or use a hosted provider like Together, Fireworks or Groq for the same models without owning hardware.

Question 6

How does Llama compare to GPT-5.5 and Claude?

Accepted Answer

Llama 3.3 70B trails GPT-5.5 and Claude 4 Opus on raw intelligence (66 vs 82 / 81) but closes most of the gap on standard chat and writing tasks. For workloads that require data privacy, self-hosting or on-premise deployment, Llama is often the better choice even with the quality difference.

Question 7

How much does Llama cost to use?

Accepted Answer

Self-hosted, Llama costs only your GPU bill — effectively free per token at high utilisation. Via hosted APIs, Llama 3.3 70B runs about $0.23 / $0.40 per 1M tokens on Together, Fireworks or Groq. Llama 3.1 8B drops to $0.06 / $0.06 — the cheapest production-grade model in our catalog.

Question 8

Which Llama model is best for coding?

Accepted Answer

Llama 3.3 70B is the strongest Llama model for coding tasks. It trails Claude 4 Opus and GPT-5.5 on HumanEval and SWE-bench, but for self-hosted code agents where you want full data control, it’s the best open-weights option short of DeepSeek V3.

Question 9

What is the “intelligence index”?

Accepted Answer

A 0–100 composite of public benchmarks (MMLU, MMLU Pro, GPQA, MATH, HumanEval). Higher is better.

Question 10

How often is this updated?

Accepted Answer

Monthly. Each model has a lastUpdated stamp. Open an issue if you spot something stale.


Llama 3.3 70B*	Meta	66	200 tok/s	0.4s	128k	$0.23	$0.40	$0.27
Llama 3.1 405B*	Meta	64	32 tok/s	0.7s	128k	$2.70	$2.70	$2.70
Llama 3.1 8B*	Meta	44	750 tok/s	0.2s	128k	$0.06	$0.06	$0.06

Meta AI Models Meta

Meta models

Browse by category

By License

Open Source

Proprietary

By Purpose

Frontier

Reasoning

Fast & Cheap

Code

Multimodal

By Modality

Text-to-Image

Text-to-Video

Text-to-Speech

By Provider

OpenAI

Anthropic

Google

Mistral

DeepSeek

xAI

Frequently asked questions

Explore the full catalog