Vendor-compiled long retrieval
WinnerMRCR 512K–1M
Meta scorecard. Claude Opus 5 is marked unavailable for this context band.
SPIRITT
Muse Spark 1.3Muse Spark 1.3 makes a 20-point DeepSWE jump
Meta's September 2 release targets long-horizon coding, browser, and desktop agents. Its launch scorecard reports 75.4% on DeepSWE v1.1, 98.1% retrieval across 512K to 1M tokens, and 20% fewer tool calls. Muse Spark 1.3 is available in the SPIRITT model picker for agentic apps and workflows.
Muse Spark 1.3 across X
Every original post remains. New official showcase posts add a rendering engine, a playable 3D kart racer, and a production-style guitar tuner, each built from a one-shot prompt.
A release built for longer, more efficient agent loops
Meta's scorecard shows a broad step up from Muse Spark 1.2. Independent testing now confirms a frontier-adjacent xhigh model, while also clarifying the limits of the max configuration.
What Meta launched, and what you can build in SPIRITT
Meta positions Muse Spark 1.3 as a major upgrade for coding, UI understanding, browser use, desktop automation, and multi-hour agentic workflows. The company reports more reliable tool use, better task planning, 20% fewer tool calls, and 25% fewer tokens than Muse Spark 1.2 in internal comparisons.
Meta's launch scorecard reports 75.4% on DeepSWE v1.1, 59.4% on SWEAtlas CodeBase QnA, 66.9% on OSWorld 2.0, and 98.1% on MRCR's 512K-to-1M-token band. Those are vendor-compiled results drawn from a mix of official leaderboards, self-reported comparator scores, and Meta-run evaluations under the linked methodology.
Artificial Analysis independently measured the available xhigh configuration at 61 on its launch-day Intelligence Index and 64 on its Coding Agent Index. The max configuration reached 62 and 68 respectively, but remained in limited partner preview and did not have public pricing at verification time.
Muse Spark 1.3 is available in the SPIRITT model picker. Select it inside a fully equipped workspace to build agentic apps and workflows with browser, terminal, files, tools, memory, and integrations.
75.4% DeepSWE61 independent Index20% fewer tool calls25% fewer tokensMax in limited previewAvailable in SPIRITT
GDPval-AA v2 Elo
Meta launch scorecard. Muse Spark 1.3 max at 1,754 improves 139 points over 1.2 and sits between the displayed GPT-5.6 Sol and Claude Opus 5 results.
Where Muse Spark 1.3 advances, ties, and still trails
All 1.3 figures below come from Meta's launch scorecard and methodology. They are not independent SPIRITT reruns.
Vendor-compiled coding agent
WinnerDeepSWE v1.1
Meta scorecard. The 75.4% result is 20.4 points above Muse Spark 1.2.
Vendor-compiled codebase knowledge
WinnerSWEAtlas CodeBase QnA
Meta scorecard. Higher is better.
Vendor-compiled terminal agents
WinnerTerminal-Bench 2.1
Meta scorecard. Muse Spark 1.3 ties the displayed GPT-5.6 Sol result.
Vendor-compiled computer use
Near frontierOSWorld 2.0
Meta scorecard. Muse Spark 1.3 nearly matches the strongest displayed result but does not lead.
Vendor-compiled job tasks
Near frontierJobBench
Meta scorecard. Muse Spark 1.3 is 0.8 points behind the displayed leader.
Vendor-compiled limitation
DeepSearchQA
Meta scorecard. Strong absolute result, but behind both displayed frontier comparators.
Vendor-compiled automation
Near frontierAutomationBench
Meta scorecard. Muse Spark 1.3 improves sharply over 1.2 but remains just behind the displayed leader.
Vendor source: Meta's September 2 Muse Spark 1.3 launch scorecard and linked multimodal evaluation methodology. Meta selects the highest comparable published score for each system from official leaderboards, self-reported results, or its own runs, so every chart remains labeled vendor-compiled. Independent context: Artificial Analysis's September 2 and 3 evaluations of the xhigh and limited max configurations. Evaluate the released configuration on your own workloads before production use.
How It Works
Choose Muse Spark 1.3 in the picker, then build the agentic workflow
01
Open a workspace
Open a workspace and land in a fully equipped cloud computer: browser, files, terminal, integrations, and memory. No local setup. No thin chat box pretending to be an agent.
02
Choose Muse Spark 1.3
Open the SPIRITT model picker and select Muse Spark 1.3. Then give it the tools and acceptance path for a real coding, browser, or long-horizon agent task.
03
Build or automate
Tell it what to ship or what to run. Muse Spark 1.3 can code, call tools, drive the browser, coordinate multi-step work, and keep going while you step away. The point is not another chat window. It is an agentic environment where Muse Spark 1.3 actually does the job.
Build with Muse Spark 1.3 now.
Open a SPIRITT Workspace, choose Muse Spark 1.3 in the model picker, and turn the release into a working agentic app or workflow.
Questions
Muse Spark 1.3 FAQ
01What is Muse Spark 1.3?+
Meta's September 2026 upgrade to its Muse Spark model for coding, browser use, desktop automation, UI understanding, and long-horizon agentic tasks.
02How much better is it than Muse Spark 1.2?+
Meta's scorecard shows broad gains, including DeepSWE rising from 55.0% to 75.4%, OSWorld 2.0 from 47.6% to 66.9%, and 512K-to-1M MRCR from 55.5% to 98.1%. Meta also reports 20% fewer tool calls on average.
03Are these independent benchmark results?+
The displayed capability cards are still Meta-compiled. Its methodology combines official leaderboard results, self-reported comparator scores, and Meta-run evaluations. Artificial Analysis has now independently evaluated both xhigh and max, reporting 61 and 62 on its launch-day Intelligence Index and 64 and 68 on its Coding Agent Index.
04Is max reasoning broadly available?+
Not at verification time. Artificial Analysis described max as a limited partner preview, while xhigh was the broadly available configuration. That matters because Meta's headline scorecard uses max and may not match the version every user can select.
05What are the main limitations?+
The vendor scorecard still trails the strongest displayed comparator on GDPval-AA v2, JobBench, OSWorld 2.0, DeepSearchQA, Agentic IF, and AutomationBench. Independent testing also shows that xhigh gains can require more input tokens per task, while max remains limited and lacks public pricing.
06Does Muse Spark 1.3 support very long tasks?+
That is a central release claim. Meta targets multi-hour workflows and reports 98.1% on its 512K-to-1M-token MRCR evaluation band. Real applications should still test retrieval and tool-state reliability across their own longest traces.
07Is Muse Spark 1.3 available in SPIRITT Workspaces?+
Yes. Open a SPIRITT Workspace, choose Muse Spark 1.3 in the model picker, and use it for coding, browser, desktop, and long-horizon agentic workflows.
Buy from builders who use what they sellBuilt using