Diffusers 0.40.0 ships MiniMax song + video-audio, Stable Audio 3
The open-weights media stack grew overnight: songs from lyrics, synced video+audio, stereo audio — all self-hostable. Plus JAX/Flax is finally gone.
Copy markdown
One release, three new media stacks
Diffusers 0.40.0 graduates Modular Diffusers out of experimental and adds tensor-parallel inference on CUDA and AWS Neuron, plus SDNQ and Nunchaku Lite quantization backends. LTX-2.5 video support lands with an updated Gemma 4 text encoder.
MiniMax Music 3: a full song from your lyrics
Feed it lyrics with [Verse]/[Chorus] tags plus a style-and-instrumentation prompt and it returns a complete song up to five minutes, as 32 kHz 16-bit stereo WAV. Open weights on the Hub, running on CUDA via Diffusers, ComfyUI, or SGLang-Omni — local backing tracks and vocals with no SaaS in the loop.
MiniMax-H3: video and its soundtrack in one pass
It denoises 5–15s clips at 24fps together with synchronized stereo audio — no separate vocoder — across three workflows: text-only, first/last keyframe, and reference (up to 12 image/video/audio clips). The catch: ~125 GB in bfloat16 before quantization, so it wants a serious GPU.
Stable Audio 3: stereo sound you can inpaint
Three pipelines ship — generation, audio-to-audio, and inpainting — for stereo audio synthesis. The inpaint path lets you regenerate just a slice of a clip instead of re-rolling the whole thing.
Check these breaking changes before you bump
JAX/Flax support is removed (~12k lines deleted) and torch_dtype is deprecated in favor of dtype. If you pin diffusers in CI, drop any Flax code paths and rename your dtype kwargs before moving to 0.40.0.
Elsewhere: OpenAI adds a prompt-caching dashboard
A new API-platform dashboard surfaces cache hit rates, reads-per-write ratios, and token breakdowns, filterable by model and service tier — so you can finally see whether your caching is actually cutting the bill.