Meta · 2024 · 12 · 06 · Model · ~1 min read
Meta released Llama 3.3 70B
What's actually new
- Almost as good as the giant. Llama 3.1 405B was Meta's huge model. The new 70B is close to it on instruction-following, reasoning, and tool use.
- Cheap to host. The 70B fits on a single high-end GPU after compression.
- Same friendly licence as earlier Llamas — usable for most commercial work, with the standard Meta terms for very large companies.
If you want more
Worth knowing
- Benchmarks are within margin of error. On some long-document tasks, the older 405B still wins outright.
- 'Open weights' isn't fully open. You can run it. You can't see the data it was trained on, and very-large-scale commercial use needs a separate Meta agreement.
Who should care
Companies running their own AI on their own servers. Privacy-sensitive teams that can't send data to OpenAI or Anthropic. Developers wanting one capable model that fits on smaller hardware.
What to do about it
If you're already running Llama 3.1 70B or 405B in production, plan a switch — same hardware, better accuracy. Starting fresh? Begin here, unless you specifically need image input (Llama 3.2 covers that) or extremely long documents.
Honest take
Llama 3.3 quietly shipped the most boring-but-important model of late 2024: open-source no longer asks you to choose between 'big and capable' and 'small enough to run'. The leading edge still belongs to closed labs like OpenAI and Anthropic. But the daily-driver AI — the one most companies actually put in production — is now firmly an open-source conversation.
Sources
- Meta — Llama 3.3 model cardvendor
- Hugging Face — Llama 3.3 70B Instructthird party
- Artificial Analysis benchmarksbenchmark
Last verified · 2026 · 05 · 05 · Found a fact wrong? corrections@aguidetocloud.com