Timeline

What’s new in open models

Releases, launches and announcements across models, hardware and tools — each linked to the things it affects.

New open-weight models and new versions of existing ones.Show everything

2026 1

5 AugModel
Qwen3.8 released
Qwen3.8 27B weights published on Hugging Face by Qwen Team (Alibaba Cloud).
Model Qwen3.8 Organization Qwen Team (Alibaba Cloud)
model release Hugging Face

29 AprModel
Qwen3 released with MoE and hybrid thinking
Includes the 30B-A3B MoE that runs quickly on consumer hardware.
Model Qwen3
model release Fixture
12 MarModel
Gemma 3 released with vision and 128K context
Sizes from 1B to 27B.
Model Gemma 3
model release Fixture
30 JanModel
Mistral Small 3 (24B) released under Apache 2.0
Latency-focused 24B model.
Model Mistral Small 3
model release Fixture
20 JanModel
DeepSeek-R1 and distilled models released under MIT
Open reasoning model plus six distilled dense models based on Qwen and Llama.
Model DeepSeek-R1 Organization DeepSeek
model release Fixture

12 DecModel
Microsoft releases Phi-4
14B model with strong math and reasoning results.
Model Phi-4
model release Fixture
6 DecModel
Llama 3.3 70B released
70B instruct model approaching 405B quality.
Model Llama 3.3
model release Fixture
12 NovModel
Qwen2.5-Coder 32B released
Code model competitive with much larger closed models on HumanEval.
Model Qwen2.5-Coder
model release Fixture
19 SeptModel
Qwen2.5 family released
Seven dense sizes plus Coder and Math specialisations.
Model Qwen2.5 Model Qwen2.5-Math
model release Fixture
15 AugModel
Hermes 3 fine-tunes of Llama 3.1 released
Nous Research released Hermes 3 with improved function calling.
Variant Hermes 3 Llama 3.1 8BModel Llama 3.1
model release Fixture
23 JulModel
Meta releases Llama 3.1 (8B, 70B, 405B)
First open-weight frontier-scale Llama with 128K context.
Model Llama 3.1 Organization Meta
model release Fixture