Prompt

Which public model checkpoints are known for open licensing terms and downloadable weights?

Latest observation

Jul 20, 2026 · OpenAI APIWeb search: off

Here are some widely used public model checkpoints that are known for open licensing terms and downloadable weights. I’ll group them by ecosystem and note the license style where commonly understood.

Large language models

  • Llama 3 / Llama 3.1 / Llama 3.2 — Meta’s “Llama license” (custom open-ish terms), weights downloadable.
  • Mistral 7B / Mixtral / Ministral — released under permissive or open model terms depending on checkpoint; weights available on Hugging Face.
  • Gemma / Gemma 2 — Google’s Gemma license, weights downloadable.
  • Falcon — many checkpoints under Apache 2.0, weights public.
  • OLMo / OLMo 2 — fully open-focused releases from AI2, weights downloadable, permissive licensing.
  • Pythia — openly released by EleutherAI, weights downloadable, permissive terms.
  • GPT-Neo / GPT-J / GPT-NeoX — open checkpoints from EleutherAI/others, downloadable weights.
  • Qwen / Qwen2 / Qwen2.5 — publicly downloadable; license varies by checkpoint but many are available under permissive/open terms.
  • BLOOM — open weights, RAIL-style responsible AI license.
  • RedPajama — openly released checkpoints, weights downloadable.

Vision / multimodal

  • CLIP / OpenCLIP — open implementations and downloadable checkpoints, usually permissive.
  • Stable Diffusion 1.x / 2.x / SDXL — downloadable weights with model-specific open/commercial-use terms depending on version.
  • SAM / SAM 2 — Meta’s Segment Anything models, weights public with model license.
  • DINOv2 — downloadable weights, permissive-ish research-friendly terms.
  • Whisper — open weights, MIT license for code and public checkpoints.

Speech / audio

  • Whisper — downloadable checkpoints, permissive.
  • wav2vec 2.0 — public checkpoints from Meta/others, weights downloadable.
  • HuBERT — public checkpoints, downloadable.

Older but still notable

  • BERT — original Google checkpoints and many derivatives, weights public.
  • RoBERTa — public checkpoints, downloadable.
  • T5 / FLAN-T5 — downloadable weights, generally permissive/open.
  • XLM-R — downloadable weights, open research use.

Where to find them

Most of these are hosted on:

  • Hugging Face Hub
  • GitHub releases / model cards
  • Official project pages from Meta, Google, AI2, Mistral, etc.

Important caveat

“Open licensing terms” varies a lot:

  • Permissive open-source-like: Apache 2.0, MIT, BSD
  • Open but not OSI-open: Llama license, Gemma license, many RAIL-style licenses
  • Research-only or usage-restricted: some checkpoints permit downloads but impose conditions

If you want, I can also give you a shortlist of the most permissively licensed checkpoints only (e.g. Apache 2.0/MIT-style), or a table with exact license names and download links.

Observation history

Unlock observation history

Create a free Obsurfable account to unlock historical responses, model comparisons, and deeper brand intelligence.

How did Obsurfable measure this prompt?

Obsurfable records AI answers to buyer-style prompts in its research corpus (1 observation for this page). Metrics are distributions over observations, not a single static ranking.

Which AI systems does Obsurfable collect answers from?

OpenAI, ChatGPT, Google, Gemini, Google AI Mode, Anthropic, Claude, Perplexity, Grok, DeepSeek, Mistral, Copilot, and Meta AI.