Muse Spark 1.3: Long-Horizon Agents with Leaner Coding Workflows

Muse Spark 1.3 is available on RouterHub as meta/muse-spark-1.3. Built for long-horizon agentic and coding work, the model combines a 1,048,576-token context window with text, image, file, audio, and video inputs and a maximum output of 131,072 tokens.

What changed in Muse Spark 1.3?

Muse Spark 1.3 focuses on the parts of agentic work that become difficult as tasks grow longer: preserving requirements, coordinating several workflows, working through conflicting inputs, and knowing when to ask for clarification or confirmation.

Meta also reports a leaner coding workflow. In internal comparisons conducted by Meta engineers, Muse Spark 1.3 used approximately 20% fewer tool calls and 25% fewer tokens than Muse Spark 1.2. These figures are provider-reported rather than RouterHub-run results, and teams should validate the effect on their own tools, prompts, and acceptance criteria.

Muse Spark 1.3 at a glance

Detail	Muse Spark 1.3
Provider	Meta
RouterHub model ID	meta/muse-spark-1.3
Availability	Available on RouterHub
Inputs	Text, image, file, audio, and video
Output	Text
Context window	1,048,576 tokens
Maximum output	131,072 tokens
Workflow focus	Long-horizon agents, multi-agent coordination, coding, and multimodal work

Why does long-horizon behavior matter?

Short prompts test whether a model can produce a plausible answer. Long-running workflows test whether it can preserve the objective while the work changes around it.

Muse Spark 1.3 is designed to maintain information across extended tasks, generate context through tool use, identify gaps in its plan, and keep track of what it has learned. It is also trained to handle interruptions and new instructions without mapping them to the wrong task.

For teams building agents, this shifts evaluation toward completed work: whether the model finishes the right task, carries constraints through every stage, and produces a deliverable that is ready for review.

Which workloads should teams test first?

Repository-scale coding

Evaluate Muse Spark 1.3 on work that spans investigation, planning, implementation, testing, and revision. Useful test cases include multi-file feature development, difficult debugging, refactoring, code review, and tasks that require repeated tool use.

Measure test pass rate, regressions, instruction adherence, failed loops, and the amount of human correction needed before the change is ready to merge.

Multi-agent workflows

The model is designed to juggle multiple workflows within a long thread. For orchestrated systems, test whether it preserves role boundaries, reconciles conflicting findings, and produces a coherent final result rather than simply combining several partial outputs.

Multimodal professional work

With text, image, file, audio, and video inputs, Muse Spark 1.3 can be evaluated on workflows that combine documents, screenshots, recordings, media, and written instructions. Test each modality independently before combining them, and use the formats and quality levels the production application will actually receive.

Tasks that require judgment about when to pause

Some workflows should not continue silently when a requirement is ambiguous or an action has consequences. Muse Spark 1.3 is designed to ask focused questions, request help when blocked, and confirm consequential steps. Teams should test whether those pauses happen at the right moments and whether the model resumes without losing earlier constraints.

How should teams evaluate Muse Spark 1.3?

Start with a bounded set of representative tasks and compare the model with the current route under the same tools, context, and scoring rules.

1. Measure completed-task quality, not output volume.
2. Track tool calls, token use, retries, latency, and human interventions for the complete task.
3. Include conflicting, incomplete, and noisy inputs alongside clean examples.
4. Test interruptions, changed requirements, and failed tool calls.
5. Review whether clarification and confirmation behavior reduces risk without creating unnecessary stops.
6. Roll out gradually and retain a fallback route until results are consistent on representative traffic.

Muse Spark 1.3 is most interesting where a task must stay coherent across many steps, inputs, and tools. The practical question is not whether every workload should move to the newest model. It is where longer context, multimodal perception, and leaner tool use improve the finished result.