Meta AI Released Muse Spark 1.3: An Agentic Coding Model That Uses ~20% Fewer Tool Calls and ~25% Fewer Tokens Than Muse Spark 1.2
MarkTechPost Asif Razzaq ● Covered by 7 sources
Meta just shipped Muse Spark 1.3, a coding-focused AI for long tasks. It uses fewer tool calls and tokens, but the model stays closed.
Based on reporting by MarkTechPost, Asif Razzaq — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Meta Superintelligence Labs has released Muse Spark 1.3, the fourth Muse Spark update in five months. This one is aimed less at one-shot text generation and more at long, messy agent work: coding, keeping track of a thread, and knowing when it’s lost.
The model is available now in Muse Code and through the Meta Model API, so developers can use it in production today. But there are two clear catches. The weights are closed, so self-hosting is off the table, and the max reasoning mode is still waiting on more safety testing.
Meta says it trained Muse Spark 1.3 across multiple agent harnesses so the behavior would carry across environments instead of getting stuck in one setup. The model is built to juggle several workflows inside a single long thread, ask clarifying questions when a prompt is vague, pull the user back in when it stalls, and check before doing something consequential. It can also run more quietly or keep sending status updates, depending on what the user wants.
On the efficiency side, Meta says Muse Spark 1.3 was trained more heavily on long-horizon coding tasks and is less chatty than Muse Spark 1.2. In internal comparisons, it used about 20% fewer tool calls and about 25% fewer tokens. That matters because fewer round trips and fewer billed tokens usually mean lower cost on agentic jobs.
The scorecard is strong, though not perfectly apples to apples. Meta’s launch numbers use the max mode, which developers can’t access yet, while the shipping version is xhigh. On its own benchmark table, Muse Spark 1.3 leads or ties several rivals on coding and retrieval tasks, including DeepSWE v1.1, Terminal-Bench 2.1, and MRCR v2, where it shows especially large gains on long-context retrieval. Pricing stays unchanged at $1.25 per million input tokens and $4.25 per million output tokens, with a $0.10/$0.20 contributor tier.
My take — AI-written commentary, not fact-checked reporting
Closed weights plus a safety-gated top mode is classic Meta: big launch energy, limited public freedom. The industry keeps calling this “agentic,” but the real tell is still cost and control, not the demo reel. Open models can be messy; closed models just make the mess someone else’s problem.
Read more about this at: MarkTechPost