AI Briefing25 Major Open-Weight Models Released
·2026.06.06 18:53
Key point
More than 25 new open-weight models have been released across various fields including LLMs, images, audio, and video.
Details
LLM (Large Language Models)
- NVIDIA Nemotron 3 Ultra: A 550B-scale hybrid Mamba-MoE model that supports 1M context and recorded MMLU 89.1.
- Google Gemma 4 12B: An any-to-any model supporting text, image, audio, and video, with 256k context and support for over 140 languages.
- StepFun Step-3.7-Flash: A 198B sparse MoE VLM.
- Liquid AI LFM2.5-8B-A1B: An edge-focused MoE model with 1.5B active parameters.
- JetBrains Mellum2-12B-A2.5B-Thinking: JetBrains' first open MoE model, with excellent coding performance.
Image Generation
- Ideogram 4: Ideogram's first open-weight model, using a 9.3B flow-matching DiT, with very strong image generation capability including text rendering.
Audio and Speech
- Boson Higgs Audio v3 4B: A TTS model supporting 102 languages and 21 emotions.
- RedNote dots.tts: A codec-free, fully continuous open TTS pipeline.
- Google Magenta RealTime 2: A real-time music generation model with sub-200ms latency.
- NVIDIA Nemotron-3.5 ASR: An ASR model offering high concurrent streaming performance.
Vision and Video
- PaddleOCR-VL-1.6: A 1B-parameter SOTA document parsing model.
- Baidu NAVA: A 6.3B-scale combined audio-video generation model.
- NVIDIA Cosmos3-Super: A 64B-scale omnimodal world model for physical AI.
- JD JoyAI-Echo: A multi-shot text-to-video generation model with up to 5-minute length.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.
Related
Stanford and NVIDIA Release 'Contrastive Language Model' CLM with 9x Faster Inference