MiniMax Drops Music 3, an Open-Weight AI That Writes 5-Minute Songs
MiniMax releases an open-weight text-to-music model that generates full 5-minute songs in 32 kHz stereo with native ComfyUI support.
20 stories
MiniMax releases an open-weight text-to-music model that generates full 5-minute songs in 32 kHz stereo with native ComfyUI support.
webAI post-trained IBM's Granite 4.2 8B into a reasoning model that tops MATH-500 at 98.2 and matches frontier code scores.
ComfyUI repackaged RelayVSR, a streaming video super-resolution model that upscales clips to 1080p and 4K using sparse keyframe references and a lightweight per-frame decoder.
A trending community LoRA for Krea 2 brings consistent Japanese character generation to both the undistilled Raw base and the fast Turbo checkpoint.
Xiaomi's new 0.8B vision-language model matches multi-billion parameter OCR systems on document parsing and visual question answering benchmarks.
Microsoft's compact BERT-based prompt compressor is quietly becoming a workhorse, reframing token pruning as classification and delivering 3-6x speedups over entropy-based methods.
Microsoft's domain-specific BERT, trained from scratch on PubMed abstracts and PubMedCentral full text, keeps dominating biomedical NLP leaderboards years after release.
Video Rebirth compressed MiniMax H3 from 32B to 14B with INT8 quantization and 6 inference steps, enabling real-time 768p video generation on 4x B200 GPUs and local inference on consumer RTX cards.
Deep Principle post-trained GLM-5.3 on chemistry lab records and discovery trajectories, releasing a 753B open-weight model that beats GPT-6-Astra on FrontierScience-Research.
Rikiya Takehi and collaborators published an edited Qwen3.5-0.8B on Hugging Face, showing GPT-6 Astra can rewrite weights directly with no training.
Pruna AI converted MiniMax-H3 video model into an open image generator that runs on 21 GB of VRAM at 5x the speed.
169Pi's 9B vision-language model runs offline on laptops and phones, hitting 61.20% on SWE-Bench Verified with Apache 2.0 weights.
OrcaRouter shipped GGUF builds of an abliterated 40B MoE coding model that run on a 16GB or 32GB GPU with 262K context.
A community retrofit grafts Qwen Sparse Attention onto Qwen3.8-27B, nearly tripling prefill speed at 256k context while matching dense quality.
A community repack of Alibaba's 8-step image generator ships in GGUF, FP8, NVFP4 and MLX formats, hitting 2.5M downloads and 4.5k likes.
A partially decensored, vision-capable GGUF build of Swift 1.5 Qwen3.8-27B that still scores 97.6% on GSM8K and runs locally in 25 GB VRAM.
A community release packs MiniMax H3's omni-modal video generation into 8.9-21.6 GB GGUF files that run on consumer GPUs through ComfyUI.
A community C++ port brings BreezeBlue's 3B bilingual TTS model to any GPU through Vulkan, with voice cloning and sub-realtime generation on consumer hardware.
Sentence Transformers, the SBERT library behind over 16,000 embedding models on the Hub, moves from UKP Lab at TU Darmstadt to Hugging Face stewardship.
Edge0 released a 0.6B text-to-speech model with zero-shot voice cloning, 23-language support, and competitive accuracy against models up to 8x larger.