Meta's Muse lineup is no longer a one-model story. Over the opening days of September 2026, the company behind the flagship Muse Spark model slipped three updates through in rapid succession: a real-time voice transcription model, a coding agent that finally left beta with a subscription attached, and its image generation model landing on a third-party platform. Here is what landed, in one sweep.
Muse Voice Transcribe brings real-time dictation to Meta AI and Muse Code
Meta's first real-time audio perception model is here, and it is not a classic speech-to-text tool. Built by Superintelligence Labs, Muse Voice Transcribe is an autoregressive multimodal model from the Muse Spark family that treats audio as a stream rather than a recording to be decoded later.
The mechanics are striking. The model ingests audio in 80-millisecond chunks, roughly 12.5 samples per second, and packs each chunk into a single soft token. From there it decides on the fly whether to keep listening or emit a text token. The trickiest part is what Meta calls delay: how much audio context the model consumes before transcribing each word. There is a real tradeoff between accuracy and latency, so the model uses adaptive delay, shifting the setting word by word based on difficulty.
- 80-millisecond audio chunks folded into single soft tokens
- Adaptive delay tuned by reinforcement learning with word-error and delay rewards
- Speaker diarization and endpointing built into the streaming core
- 70+ languages trained, with 25 extensively validated on release
That adaptive behavior is shaped through reinforcement learning that balances a word error rate reward against a delay reward. Meta says the result sits on the Pareto front of the speed-accuracy tradeoff, measured by time to final transcription. It also layers speaker diarization and endpointing directly onto the streaming core, predicting a token when the speaker switches and tagging who is talking at the end of each chunk.
The language coverage is ambitious: 70-plus languages in training, with 25 comprehensively verified for the initial rollout, and the model supports arbitrary code-switching, even swapping languages within a single sentence. In practice, Muse Voice Transcribe already powers voice dictation across Meta AI and Muse Code, usable in any application or window with a press-and-hold key to activate.
Muse Code leaves beta with a price tag
Meta's first coding agent went live in beta in early August as a terminal tool that plans changes, writes code, and validates results across large repositories. It fans big jobs out to parallel sub-agents working in isolated worktrees, so the working copy is never touched, and it is powered by Muse Spark 1.2, a coding-focused model update.
Now it graduates out of beta. Reports point to paid plans ranging from roughly $5 to $50, plus a developer preview of the SDK going public. The move is part of Meta's broader push to actually monetize its models. Recent reporting on Muse Spark noted per-token fees running at roughly 25 percent of rival pricing, and the agent's cheapest contributor tier remains more than 10 times cheaper than competitors, at the cost of letting Meta see usage data to keep improving the model. Customers who want privacy can pay more to keep their data out of Meta's hands.
Muse Image finds a home on fal
The image side of the family is expanding beyond Meta's own walls. Muse Image, the agentic image model from the same lineup, is now available on fal, the generative media platform favored by developers building agentic pipelines. That gives teams a way to call Meta's image generation as part of larger automated media workflows, rather than only through Meta's own stack.
The bigger picture is that Meta is widening the Muse moat beyond a single flagship. By turning audio, code, and image into one fast-moving product family, Superintelligence Labs is signaling that its revival is a release train, not a one-off drop.
Voice at the front, a subscription on the coding side, and third-party reach for images all add up to a clear strategy: make the Muse brand inescapable across every generative format at once.
Comments