Params

800M

The original GPT-2 checkpoint. Still the most downloaded causal language model on the Hub and a standard baseline.

Downloads 30d

14.3M

Hub likes

3.5K

Params

124M

Mid-size dense model from the Qwen3 family, with switchable thinking and non-thinking inference modes.

Downloads 30d

13.6M

Hub likes

1.3K

Params

8B

Params

31B

Instruction-tuned 7B chat model with strong multilingual coverage and a permissive licence.

Downloads 30d

10.6M

Hub likes

1.6K

Params

7B

Params

19B

Params

500M

Smallest checkpoint in Meta’s Open Pre-trained Transformer suite, released to reproduce GPT-3 class training openly.

Downloads 30d

11.1M

Hub likes

294

Params

125M

Params

31B

Small instruction-tuned model that runs comfortably on modest hardware.

Downloads 30d

7.6M

Hub likes

811

Params

1.5B

Small Qwen3 model supporting both thinking and direct-answer modes in a single checkpoint.

Downloads 30d

5.8M

Hub likes

687

Params

4B

Compact instruction-tuned model built for on-device assistants and summarisation at low latency.

Downloads 30d

6.6M

Hub likes

1.6K

Params

1B

OpenAI’s smaller open-weight mixture-of-experts release, roughly 21B total parameters with about 3.6B active per token.

Downloads 30d

6.5M

Hub likes

5K

Params

21B

The most-liked chat model on the Hub. 128K context, instruction-tuned, widely used as a fine-tuning base.

Downloads 30d

5.9M

Hub likes

6.7K

Params

8B

Params

3B

Params

9B

The larger open-weight OpenAI release, around 117B total parameters with roughly 5.1B active per token.

Downloads 30d

5.3M

Hub likes

5.1K

Params

117B

Largest dense model in the Qwen3 line, aimed at reasoning and long-context work under a permissive licence.

Downloads 30d

4.8M

Hub likes

740

Params

32B

Params

34B

Params

304B

Params

35B

Params

4B

Params

72B

Lightweight Qwen3 checkpoint aimed at latency-sensitive chat deployments.

Downloads 30d

3.8M

Hub likes

526

Params

1.7B

Params

7B

Params

9B

Part of a model suite trained on identical data in identical order, built specifically for interpretability research.

Downloads 30d

3.5M

Hub likes

45

Params

160M

Params

4B

Instruction-tuned small model from the Gemma 3 family, built from the same research as Gemini.

Downloads 30d

3.4M

Hub likes

1.1K

Params

1B

Code-specialised instruction model covering generation, completion and repair across many languages.

Downloads 30d

2.7M

Hub likes

800

Params

7B

Params

27B

Params

14B

Mid-size instruction-tuned chat model, a common step up when 7B is not quite capable enough.

Downloads 30d

2.7M

Hub likes

362

Params

14B

Params

33B

Params

685B

Params

35B

Very small language model trained on a curated corpus, intended for on-device and edge deployment.

Downloads 30d

2.5M

Hub likes

229

Params

135M

Params

35B

Params

88M

Params

14B

Params

27B

Params

31B

Params

26B

Mid-size member of the Qwen2.5-Coder family, trained on a large code-heavy corpus.

Downloads 30d

1.8M

Hub likes

188

Params

14B

Mixture-of-experts model with roughly 30B total parameters and about 3B active, trading memory for inference speed.

Downloads 30d

2.4M

Hub likes

930

Params

30B

Params

3B

Params

7B

Params

8B

Params

31B

Params

500M

Params

14B

Params

1.7B

Params

4B

Params

32B