September 2, 2026
Meta released Muse Spark 1.3 on September 2 and made it available, with maximum reasoning, in its Muse Code coding agent and the Meta Model API, the company said. Meta said the model “delivers improved performance across agentic and coding tasks.”
Takeaway points
- In comparisons by Meta engineers, Muse Spark 1.3 used about 20% fewer tool calls and 25% fewer tokens than Muse Spark 1.2, the company said.
- Meta said the model now asks clarifying questions when prompts are ambiguous and confirms before it takes consequential actions.
- The roadmap in Meta’s announcement includes “the Muse Spark open weights release,” with no date given.
What is new in Muse Spark 1.3
Meta said it built Muse Spark 1.3 using what it learned “from months of broad adoption of Muse Code and Meta Model API,” and made the model easier to use in real-world settings. The company described the release as a step in its work toward what it calls personal superintelligence.
On coding, Meta said the new version was trained on more long-horizon coding tasks. Relative to Muse Spark 1.2, it takes fewer turns where they are not needed, is less verbose and has a cleaner overall coding style, according to the announcement. In comparisons run by Meta engineers, it used about 20% fewer tool calls and about 25% fewer tokens. Meta did not publish the tasks behind those comparisons in the announcement. Its examples include a draft flow-simulation report for an experimental aircraft wing assembly, built from simulation results and a CAD model.
How Meta trained Muse Spark 1.3 for longer tasks
Meta said Muse Spark 1.3 is designed to sustain longer work and to juggle several workflows in a single thread. Given an open-ended objective, the model uses tools to gather its own context from conflicting sources, corrects gaps in its plan, and tracks what it has learned before producing a final deliverable, the company said. Meta said it trained the model “across a diverse set of harnesses to generalize to various agentic environments.”
The company also described changes to how the model works with people. It asks clarifying questions when prompts are ambiguous, asks for help when stuck, and confirms before taking consequential actions. On long tasks it can post frequent updates or work silently in the background, depending on the user’s preference. Meta said it follows long, complex instructions more reliably than earlier Muse Spark models, and maps incoming prompts to the right task more accurately when a single conversation carries several.
Meta said the model has a better sense of its own limits, and flags when it hits hurdles “instead of hallucinating outcomes.” On safety, the company reported improved resistance to adversarial inputs and prompt injections, and better calibration on which actions are irreversible.
Muse Spark 1.1 and Meta Superintelligence Labs
Muse Spark 1.3 follows Muse Spark 1.1, which Meta released on July 9. Meta described 1.1 as the latest model from Meta Superintelligence Labs and “a multimodal reasoning model built for agentic tasks,” with a context window of 1 million tokens. Alongside 1.1, Meta launched a public preview of the Meta Model API, its first paid route for developers to build on the model, and made the model available in “Thinking” mode in the Meta AI app and on meta.ai.
Meta said 1.1 could act as a main agent that gathers context, makes a plan and hands execution to parallel subagents, and that it manages its own context window by compacting what it keeps for later steps. For computer use, the company said it trained 1.1 to write scripts when automation is faster and to click when direct interaction is simpler. The 1.1 launch came in the same week as Muse Image, a separate Meta release. One demonstration in the 1.1 announcement showed the model turning smartphone video of an item into a Facebook Marketplace listing by operating a browser, and Meta said it has “even more capable models in training.”
Meta said 1.1 was evaluated under its Advanced AI Scaling Framework across chemical and biological, cybersecurity and loss-of-control risks, and that the results fell within safe margins. The company also reported stronger resistance to jailbreaks and to prompt injection from untrusted data, lower hallucination rates and less sycophancy. The 1.1 announcement showed the model working inside OpenCode, an open-source coding agent. It also quoted partners. Amjad Masad, chief executive of Replit, listed a million-token context, multimodal input, built-in search with citations and parallel tool calling “in a clean OpenAI-compatible package.” Saoud Rizwan, chief executive of Cline, said the model paired strong tool use with a price point that made real coding workloads viable at scale. Yashodha Bhavnani of Box said that on Box’s enterprise evaluation set it was competitive with leading frontier models. Dave Morin of the OpenClaw Foundation said it worked well for running agents with OpenClaw.
The 1.3 announcement closes with Meta’s plans: “bigger models, the Muse Spark open weights release, and more.” Meta has not said when the open weights will arrive or which version they will cover.
Which agents can run Muse Spark 1.3
Third-party agents already reach Meta’s models. Nous Research added the Meta Model API, listed with Muse Spark, as one of six new providers in Hermes Agent 0.21.0 on August 31, according to its release notes. OpenClaw’s 2026.8.1 release made a Meta provider available as a separate official plugin.
Muse Code, Meta’s own coding agent, installs on macOS or Linux through a shell script from Meta’s developer site, according to the announcement. StrideNote’s guide to coding agent file permissions covers what such an agent can reach once installed.
The release landed between two rival launches: Anthropic’s Claude Fable 5.1 the day before and OpenAI’s GPT-6 Astra the day after.
Muse Spark 1.3 pricing and open weights: what Meta has not published
Meta’s announcement gives no prices, no context window and no parameter count for Muse Spark 1.3. It does not say whether the 1-million-token context of version 1.1 carries over, and it gives no date for the open weights release.
Sources: Meta, Muse Spark 1.3; Meta, Muse Spark 1.1; Hermes Agent 0.21.0 release notes; OpenClaw 2026.8.1 release notes.