Introducing Muse Spark 1.1

TL;DR

Meta Superintelligence Labs announces Muse Spark 1.1, an upgrade to the original Muse Spark (Muse Spark — first model from Meta Superintelligence Labs (MSL)) positioned as a multimodal reasoning model built for agentic tasks — coding, computer use, and tool orchestration — with a 1M-token context window that the model “actively manages.” Distribution moves from a private preview to a public-preview Meta Model API, with a “Thinking” mode in the Meta AI app. Coding, agentic (multi-agent orchestration and delegation), and computer-use are the pitched headline capabilities; a separate evaluation report is linked but not included, and the blog itself discloses no architecture, parameters, training data, or benchmark numbers.

Key claims

Muse Spark 1.1 is an upgrade over Muse Spark, positioned on the performance-efficiency frontier, and shipped alongside the Muse Image / Muse Video launches from MSL [§intro].
The model orchestrates multi-agent systems to optimize end-to-end latency: as main agent it plans and delegates to parallel subagents; as subagent it stays scoped and escalates back when needed [§Agents].
It zero-shot generalizes to new native tools, MCP servers, and custom skills [§Agents].
Context window is 1M tokens and the model is trained to “actively manage” it — remembering earlier actions, retrieving info from earlier work, and compacting critical steps for later reuse [§Agents].
For computer use, the model is trained to choose between scripting (when automation is faster), direct clicks (when interaction is simpler), and batched action generation at each step, rather than reasoning through every desktop step one click at a time [§Computer Use].
Coding improved substantially on real-world tasks: diagnosing/fixing complex bugs, implementing features in enterprise codebases, executing large code migrations; competitive with leading alternatives on Meta Internal Coding Bench [§Coding].
Coding harness support includes planning mode, goal conditioning, subagent delegation, and context compaction — designed to plug into popular agentic coding setups [§Coding].
Multimodal capabilities include visual-to-code artifact generation, ultra-descriptive image and video captioning, and grounded outputs for computer-use workflows [§Multimodal].
Safety evaluations follow Meta’s Advanced AI Scaling Framework across Chemical & Biological, Cybersecurity, and Loss of Control risk categories; the blog claims strong resistance to direct jailbreaks, prompt injection, and developer-prompt attacks, plus reduced hallucination and sycophancy vs the prior Muse Spark [§Safety].
Public preview of the Meta Model API is now open; the model is available in “Thinking” mode in the Meta AI app [§Availability].

Method

No architecture, parameter count, training data, training compute, or benchmark numbers are disclosed in the blog. The only operational details are (a) the multi-agent training setup — the model is trained to act as both main agent and subagent with explicit delegation/escalation behavior, (b) the computer-use training to co-train scripted automation with direct GUI interaction and choose between them, (c) active context management as a trained behavior over a 1M-token window rather than a passive property of positional encoding, and (d) support for common agentic coding harness features (planning mode, goal conditioning, subagent delegation, context compaction). Separate evaluation methodology and evaluation report documents are linked but not summarized in the launch post.

Results

No numeric benchmarks are reported in the blog. Claims are qualitative: “significantly improves upon Muse Spark and is competitive with leading alternatives” on Meta’s internal coding bench, and third-party partner quotes (Replit, Cline, Box) frame the model as viable for enterprise-scale agentic coding workloads. Demo videos (dinner-party organization, OpenCode debugging, DeepSWE self-evaluation, Facebook Marketplace listing) illustrate qualitative capability but are not tied to quantitative results in the post.

Why it’s interesting

This is the direct successor to Muse Spark — first model from Meta Superintelligence Labs (MSL) (the original Muse Spark launch) and completes the MSL lineup that already includes Introducing Muse Image: Image Generation Built for Your World and the previewed Muse Video (Muse Video preview — MSL video generation model announcement (Alexandr Wang)). The pitch has shifted since Muse Spark: the first launch emphasized “contemplating mode” (parallel-agent reasoning like Gemini Deep Think), but 1.1 leans into agentic-coding and computer-use — the same product slot occupied by Introducing Composer 2.5, Kimi K2.6: Advancing Open-Source Coding, and Introducing GPT-5.3-Codex. That the model is trained to choose between scripting and direct GUI clicks is a concrete instance of the “learned harness discipline” theme Marco DeepResearch: Unlocking Efficient Deep Research Agents via Verification-Centric Design flagged — verification and action-batching baked into the model rather than deferred to a wrapper. Notably, MSL is still closed here: Muse Spark 1.1 lands as API preview only, sitting in the closed cohort of Open foundation-model releases alongside GPT-5.3-Codex and Composer 2.5, in contrast to open agentic-coding releases like Kimi K2.6 and MiniMax M2.