Meta Superintelligence Labs releases a new product this week. Muse Spark 1.3. It’s the fourth Muse Spark update in the last five months. Instead of single-turn production, the focus will be on agentic and code work with a long-horizon. Meta’s blog post is framed around usability. It is about sustaining long threads, working with users, and knowing how to fix it when you get stuck.
Is it deployable? Muse Spark 1.3 is now available. Muse Spark 1.3 is now available. Muse Code Then, there is the Meta Model APIYou can now call the project in production. It is not possible to self-host, as the weights and scales have been closed. Max. The reasoning mode remains a gated feature, awaiting further safety tests.
What has changed?
Meta-trained Muse Spark 1.3.3 across multiple harnesses of agents so that behavior can be generalized beyond one environment. This model was designed to contain multiple workflows in a long thread. The model gathers information from multiple sources to create its own contextualization.
Collaboration changes are more useful. Muse Spark asks clarification questions for ambiguous instructions, draws the user into it when it stops, and confirms any actions that may be consequential. On longer runs, it can adapt to the user’s preference for frequent status updates or quiet background execution. Meta can also report better calibrations on the model itself, so that it does not hallucinate an outcome.
The multitasking has also been improved. Meta states that its model maps an inbound prompt to the appropriate task, regardless of whether the user interrupts or steers.
Coding and efficiency
Muse Spark was taught to code for longer-term goals. Meta reports fewer unnecessary turn, less verbosity and cleaner code style compared to Muse Spark. Meta engineers found that it was using approximately 20% less tool calls, and 25% fewer tokens. This number is the cost for agentic workloads: less round-trips and fewer tokens billed per task.
‘+b.name+’
‘+
‘
‘
xhigh‘+b.fmt(b.ship)+’
‘+
‘
Max.‘+b.fmt(b.chart)+’
‘+
‘
‘+
‘

