Muse Spark 1.3
A purported Meta model circulating in community posts that has not been officially confirmed. The name suggests creative generation and fast inference capabilities, and it may be an internal code name or an unofficial designation.
Core Facts
Timeline (last 90 days)
When acting as the executor, Meta Muse Spark 1.3 tried a nonexistent command-line argument, triggering a runaway task that regenerated all 442 defined actions.
Meta Muse Spark 1.3 scored 69% when tested as a visual judge, and is suspected of having memorized the examples rather than learned the rules.
Meta divides test tasks into In-distribution tasks, which were a post-training optimization focus, and Out-of-distribution tasks, which received no targeted optimization, and compares how Muse Spark 1.3 and Muse Wonder Full perform under different reasoning budgets.
Running `/models` in OpenCode brings up the Select Model list, and in the video demonstration the list contained 9 free models: Big Pickle, Space Bunny, Cat Jam, Longcat 2.5 Preview, Step 5 Preview, Mimo v2.6 Flash, Muse Spark 1.3, and 2 Nvidia Nemotron models.
Muse Spark 1.3 has an independent 3000/4M quota pool, and its price is only one-third of Wonder's
Muse Spark 1.3 uses the same 3000 RPM and 4M TPM figures, but Spark's Contributor variant has lower limits, only 150 RPM and 3M TPM.
Developer Pavel Hurin tested Argon with a bug hunter task, and the result was comparable to Muse Spark 1.3, clearly lagging behind Astra, Fable and Opus.
Muse Wonder costs $4.52 per task call, an increase of nearly 3 times compared with Muse Spark 1.3, about $1 more expensive than GPT-6 Astra or Grok 4.7, but about $1.5 cheaper than Opus 5.5.
Muse Spark 1.3 scores 48 on the Artificial Analysis Intelligence Index, just 6 weeks after Muse Wonder's release, an improvement of 4 index points in six weeks.
Claude Code and Codex can be run on Muse Spark 1.3 via the Meta Model API.
23 more timeline events
All Facts (20)
Muse Wonder costs $4.52 per task call, an increase of nearly 3 times compared with Muse Spark 1.3, about $1 more expensive than GPT-6 Astra or Grok 4.7, but about $1.5 cheaper than Opus 5.5.
70%UnverifiedMuse Spark 1.3 uses the same 3000 RPM and 4M TPM figures, but Spark's Contributor variant has lower limits, only 150 RPM and 3M TPM.
60%UnverifiedMuse Spark 1.3 achieved an accuracy rate of 98.1% in ultra-long context tests of one million tokens.
60%UnverifiedMuse Spark 1.3 adopts a reasoning-first architecture, reducing tool calls by approximately 20% compared to previous large models.
60%UnverifiedMeta Muse Spark 1.3 scored 69% when tested as a visual judge, and is suspected of having memorized the examples rather than learned the rules.
50%UnverifiedMeta divides test tasks into In-distribution tasks, which were a post-training optimization focus, and Out-of-distribution tasks, which received no targeted optimization, and compares how Muse Spark 1.3 and Muse Wonder Full perform under different reasoning budgets.
50%UnverifiedRunning `/models` in OpenCode brings up the Select Model list, and in the video demonstration the list contained 9 free models: Big Pickle, Space Bunny, Cat Jam, Longcat 2.5 Preview, Step 5 Preview, Mimo v2.6 Flash, Muse Spark 1.3, and 2 Nvidia Nemotron models.
50%UnverifiedMuse Spark 1.3 has an independent 3000/4M quota pool, and its price is only one-third of Wonder's
50%UnverifiedDeveloper Pavel Hurin tested Argon with a bug hunter task, and the result was comparable to Muse Spark 1.3, clearly lagging behind Astra, Fable and Opus.
50%UnverifiedMuse Spark 1.3 scores 48 on the Artificial Analysis Intelligence Index, just 6 weeks after Muse Wonder's release, an improvement of 4 index points in six weeks.
50%UnverifiedMuse Spark 1.3 is the core inference model in the Meta Muse ecosystem, with code comprehension and generation capabilities.
50%UnverifiedClaude Code and Codex can be run on Muse Spark 1.3 via the Meta Model API.
50%UnverifiedDevelopers are advised to first trial the new model on non-critical tasks, observing its code generation quality, context retention capability, and compatibility with existing features, before deciding whether to adopt it long-term.
50%UnverifiedMuse Spark 1.3 achieved a score of 88.8 on the Terminal-Bench 2.1 test
50%UnverifiedMuse Spark 1.3 claims to match or surpass the top-tier code writing and long-text reasoning capabilities of Claude Opus 5 and GPT 5.6 SO.
50%UnverifiedMeta has never officially committed to open-sourcing (open weights) the Muse Spark 1.3 version, and the open-source claims circulating online are a misreading.
50%UnverifiedMuse Spark 1.3 has a first-token output latency of up to 27.5 seconds in its highest reasoning mode.
50%UnverifiedDue to high latency, Muse Spark 1.3 is unsuitable for IDE inline code completion scenarios requiring real-time feedback, and is better suited as an offline batch-processing programming assistant.
50%UnverifiedMuse Spark 1.3 has blind spots in understanding three-dimensional physical space, reversing the orientation of the chassis and screw holes when generating CAD model code for a 3D-printed sports car.
50%UnverifiedTasks involving physical space modeling, mechanical design, and 3D simulation using Muse Spark 1.3 require additional manual verification.
50%