Llama 3.1

tl;dr: Meta’s open LLM family that introduced 128K context and tool use to the open-weights tier. The 8B is a baseline assistant, the 70B is the sweet spot for most hardware, and the 405B was the first openly available model Meta benchmarked against GPT-4 and Claude 3.5 Sonnet.

What it is

llama3.1 is the umbrella tag Ollama ships for Meta’s Llama 3.1 release. Three sizes, all text, all 128K context, all with the same chat template and tool-calling prompt format. Sizes are quantized on the Ollama side — the 8B fits in 4.9GB, the 70B in 43GB, and the 405B in 243GB.

The page on ollama.com lists tools as a capability and 93 tags. Total pulls across the family are 119.3M, the highest in the Ollama library.

Pull command

From the ollama.com detail page:

ollama run llama3.1      # defaults to 8B
ollama run llama3.1:70b
ollama run llama3.1:405b

Each size pulls the latest quantized build from the Ollama registry and drops into a local REPL. The same tag names route through ollama launch claude --model llama3.1, ollama launch opencode --model llama3.1, ollama launch hermes --model llama3.1, and ollama launch openclaw --model llama3.1 — those four agents are listed on the detail page as direct integrations.

Sizes and hardware

Context is 128K on every size. Input is text-only — no vision, no audio.

What the README says

Meta’s README (mirrored on the Ollama page) describes Llama 3.1 405B as “the first openly available model that rivals the top AI models when it comes to state-of-the-art capabilities in general knowledge, steerability, math, tool use, and multilingual translation.” Meta evaluated across 150+ benchmarks and ran human evals comparing Llama 3.1 to GPT-4, GPT-4o, and Claude 3.5 Sonnet. The 8B and 70B were upgraded from the Llama 3 line with the same 128K context and tool-use support.

The README also flags a license change: Meta now allows the outputs of Llama models (including 405B) to be used to improve other models, which is what enabled the DeepSeek-R1 Llama-8B and Llama-70B distillates. The exact community license text is not surfaced on the Ollama page — pull from https://ai.meta.com/blog/meta-llama-3-1/ if you need the legal copy.

Gotchas

• The Ollama page does not list a license. The README only references “Meta’s license” generically. Don’t ship anything commercial off this page alone.
• llama3.1:8b is marked latest on the page but :latest and :8b are now separate tags — the older :latest alias still points at the 8B build, so most ollama run llama3.1 users get the small model without realizing it.
• 405B is a beast. Don’t try to load it on a workstation without checking VRAM first — Ollama will silently swap and crawl.
• The 128K context is real, but expect long-context throughput to drop as KV cache fills.

Related: DeepSeek-R1 for the open reasoning distillation story, Ollama for the runtime that serves these tags, llama.cpp for the engine underneath.

Untested here — facts from the ollama.com /library/llama3.1 page fetched 2026-09-10.