SGLang-Omni

Headlines · 3

1. 6d
   MiniMax Music 3 Brings Open-Weight Five-Minute Song Generation

   MiniMax released MiniMax Music 3, an open-weight text-to-music model that accepts lyrics and detailed musical descriptions to generate complete songs of up to five minutes, according to the MiniMaxAI model card. The architecture combines an 8B Global LLM, a 0.6B Local LLM, and continuous hidden-state synthesis using Flow Matching and Flow-VAE to produce 32 kHz, 16-bit stereo WAV audio. The model is available through Diffusers and supports SGLang-Omni inference.

   MiniMax · Hugging Face · 🔥 0
2. 6d
   INT21 says AI agents built a Drama 3 Rust runtime in one hour

   INT21 says its SwarmOS orchestrated AI agents that autonomously wrote, tested and optimized a Rust inference runtime for Fish Audio’s Drama 3 in about an hour. Its workflow generates multiple candidate implementations, tests them on hardware including NVIDIA B200 GPUs, retains working versions and discards failed ones. INT21’s benchmark reported time to first audio of 15.11–23.58 milliseconds, compared with 37.79–38.61 milliseconds for SGLang-Omni. Each test used one speech stream, a one-frame first chunk and 20 warm trials per prompt across three prompts; the results are INT21’s own reported figures.

   Wccftech · 🔥 0
3. Oct 12
   Fish Audio releases open-weight Drama 3 speech model with 35B MoE backbone

   Fish Audio has published a technical report for Fish Audio Drama 3, an open-source speech generation model that it says turns text into directed performances. The model uses a 35B-parameter mixture-of-experts backbone activating 3B parameters per token, clones voices from multiple references, follows natural-language direction, and adds multi-turn revision, multi-character scene dubbing, editing by infilling and streaming text input. The report pairs a hybrid linear-attention backbone with a new codebook multi-token-prediction audio decoder, and describes training via three-stage pre-training, instruction fine-tuning and large-batch reinforcement learning rewarded by a single learned human-preference model. Weights are said to be available on GitHub and Hugging Face with inference through SGLang-Omni, and the authors report 271 frames per second for single-stream decoding on one H200 versus 177 for Fish Audio S2.

   Hugging Face · Papers · 🔥 0

Experience and discussion from the community

Share my SGLang-Omni experienceAsk about SGLang-Omni