Quick answer
Meta released Muse Spark 1.3 on September 2, rolling it out in Muse Code and Meta Model API. The update targets long-running agentic and coding work, including managing multiple workflows in one thread, asking for clarification or help, and confirming consequential actions. Meta's internal engineering comparisons report about 20% fewer tool calls and 25% fewer tokens than Muse Spark 1.2, but the company does not publish enough detail to treat those figures as an independent benchmark. Max reasoning is still undergoing safety testing, and open weights are planned without a release date. Muse Spark 1.3 is not currently verified in Chat AI's model directory.
Muse Spark 1.3 is rolling out in Muse Code and Meta Model API
Meta says Muse Spark 1.3 became available on September 2 through Muse Code and its first-party Model API. Reasoning modes that were already offered for the model family are included at launch. The higher max-reasoning mode is not part of the broad launch; Meta says it will arrive after additional safety testing, without naming a date.
Sources: Meta AI Research, Meta AI
The update focuses on longer agent workflows and clearer handoffs
Meta describes the model as better able to maintain an open-ended objective, gather context from conflicting sources, correct gaps in its plan, and track multiple workflows inside one long conversation. The release also emphasizes collaboration behavior: the model is trained to ask clarifying questions when a request is ambiguous, ask the user for help when stuck, and confirm before consequential actions. These are provider claims about intended behavior, not guarantees for every harness or task.
Sources: Meta AI Research
Meta reports fewer tool calls and tokens in its own coding comparisons
In comparisons run by Meta engineers against Muse Spark 1.2, the company says version 1.3 used about 20% fewer tool calls and 25% fewer tokens while producing cleaner code and taking fewer unnecessary turns. Meta's launch post does not disclose the task count, prompt set, variance, or a reproducible harness for this specific efficiency comparison. The percentages therefore describe Meta's internal test rather than a dependable reduction in latency, API spend, or tool usage for every repository.
Sources: Meta AI Research
Meta's published evaluation uses max reasoning, which is not broadly available yet
The accompanying methodology says Meta generated Muse Spark 1.3 results through Meta Model API using max reasoning. It also says third-party models were tested on a best-effort basis and may not reflect provider-optimized performance, while reported scores can come from Meta's run, an official leaderboard, or a model provider's self-report. Because the launch post says max reasoning is still completing safety testing, readers should not assume every published Muse Spark 1.3 result represents the configuration available to typical API or Muse Code users on launch day.
Sources: Meta AI Research, Meta AI Research
Independent testing shows gains and regressions rather than a clean sweep
Artificial Analysis reports that the available xhigh variant scored 61 on its Intelligence Index, up from 57 for Muse Spark 1.2. Its evaluation also found a four-point gain on GPQA Diamond, a five-point gain on Terminal-Bench 2.1, and a 12-point gain on Tau3-Bench Banking. However, long-context retrieval on AA-LCR fell from 83% to 79%, and AA-Omniscience accuracy fell from 45% to 42%. Those mixed results reinforce the need to test the model on the exact tasks a team cares about rather than relying on one aggregate score.
Sources: Artificial Analysis
Safety improvements are announced, while open weights remain future work
Meta says Muse Spark 1.3 has stronger resistance to adversarial inputs and prompt injection and better calibration around irreversible actions. The announcement does not provide a standalone safety report for version 1.3, so these claims should be treated as Meta's assessment. Meta also lists a Muse Spark open-weights release on its roadmap, but the version 1.3 post does not say that weights are available now or provide a release date.
Sources: Meta AI Research, Meta AI
Evaluate the available mode on representative, instrumented tasks
A useful trial should run the same real repository tasks on Muse Spark 1.2 and 1.3 with equivalent tools, prompts, permission rules, and stopping conditions. Record successful task completion, total input and output tokens, tool calls, retries, elapsed time, and whether the model requests clarification or confirmation at the right moments. Test long-context retrieval separately from coding because the independent results move in different directions. Revisit any conclusion when max reasoning becomes generally available, since Meta's published max-mode results are not a substitute for measuring the launch configuration.
Sources: Meta AI Research, Artificial Analysis
Muse Spark 1.3 availability in Chat AI is unverified
The exact Muse Spark 1.3 model is not listed in Chat AI's current public model directory, so this article does not claim that it is available in Chat AI. Users can consult the directory for the models currently verified by Chat AI and treat any future Muse Spark listing as a separate availability update.
Sources: Chat AI
Frequently asked questions
What readers usually ask
Where is Muse Spark 1.3 available?
Meta says it is rolling out in Muse Code and Meta Model API. That does not establish availability through other apps or model providers.
Is max reasoning available for Muse Spark 1.3?
Not broadly at launch. Meta says max reasoning will follow after additional safety testing and does not give a release date.
Are Muse Spark 1.3 open weights available?
No open-weight release is announced as available in the launch post. Meta lists Muse Spark open weights on its roadmap but provides no timing.
Does Muse Spark 1.3 always use 25% fewer tokens?
No such universal claim is supported. Meta reports about 25% fewer tokens in internal engineering comparisons with Muse Spark 1.2, but does not publish enough detail to generalize that result to every task.
Do the launch benchmarks represent the generally available mode?
Not all of them. Meta's methodology says its Muse Spark 1.3 results use max reasoning, while Meta says that mode is still undergoing additional safety testing.
Is Muse Spark 1.3 available in Chat AI?
Its availability is unverified. The exact model does not appear in Chat AI's current public model directory.
Evidence
Sources
- Introducing Muse Spark 1.3Meta AI Research · Primary source
- Muse Spark 1.3 evaluation methodologyMeta AI Research · Primary source
- Introducing the Muse Spark model familyMeta AI · Primary source
- Muse Spark 1.3 independent evaluationArtificial Analysis · Secondary source
- Chat AI model directoryChat AI · Primary source