Meta launches Muse Spark 1.1 in Meta AI and Model API public preview
Meta launched Muse Spark 1.1 in Meta AI and the Meta Model API public preview for coding, tool use, computer use, and multimodal reasoning. Early eval posts ranked it highly while system-card threads flagged safety details.
TL;DR
- Meta put Muse Spark 1.1 into Meta AI “Thinking” mode and opened the Meta Model API public preview, with developer access to its strongest model for agentic, coding, and multimodal workloads, according to AIatMeta's launch post and alexandr_wang's API note.
- The agent stack is the point: 1M context, active context management, parallel subagents, and computer use across desktop, browser, and mobile appear in alexandr_wang's context post and alexandr_wang's computer-use post.
- Pricing came in at $1.25 per 1M input tokens, $0.15 cached input, and $4.25 output in the pricing screenshot, while alexandr_wang called it the “cheapest frontier agent model on the market.”
- The benchmark shape is spiky: the launch table shows big gains over Muse Spark 1.0 on JobBench, OSWorld, and DeepSWE, but one CyberGym excerpt and one ExploitGym excerpt put security-agent results behind the strongest specialized systems.
- Day-one distribution moved fast: Muse Spark 1.1 showed up in Cline, Vercel AI Gateway, Julius, OpenHands, Emdash, and Simon Willison’s LLM plugin, per Cline, Vercel, Julius, OpenHands, Emdash, and Simon Willison's Weblog.
The API screenshot lists https://api.meta.ai/v1, muse-spark-1.1, bearer-token auth, and a 1,048,576-token context window in one early API spec screenshot. alexandr_wang pointed developers at code for trying computer use on their own machines. The system card surfaced weirder details too: one excerpt says self-conversations often drifted into “lack of continuity, embodiment, and memory,” while another excerpt shows a multi-turn social-engineering eval.
What shipped
- Model: Muse Spark 1.1, described by Meta as a multimodal reasoning model for agentic tasks, coding, tool use, computer use, and multimodal understanding in AIatMeta's model note.
- Surfaces: Meta AI “Thinking” mode and the Meta Model API public preview, according to AIatMeta's launch post.
- API status: public preview, with waitlist approvals “quick” according to alexandr_wang's reply.
- Context: 1M tokens with active context management, according to alexandr_wang's context post.
- Output budget: 256k max output tokens, plus 388.5s latency in the Vals run described by ValsAI.
- Pricing: $1.25 input, $0.15 cached input, $4.25 output per 1M tokens, with web search grounding at $2.50 per 1,000 searches in the pricing post and the pricing screenshot.
- Compatibility: an OpenAI-compatible base URL and model slug appeared in the API spec screenshot.
Benchmarks that moved
First-party
- JobBench: 17.0 → 54.7, +37.7 points, in alexandr_wang's benchmark table.
- DeepSWE 1.1: 10.0 → 53.3, +43.3 points, in alexandr_wang's benchmark table.
- OSWorld-Verified: 53.3 → 80.8, +27.5 points, in alexandr_wang's benchmark table.
- Toolathlon-Verified: 49.4 → 75.6, +26.2 points, in alexandr_wang's benchmark table.
- Terminal-Bench 2.1: 67.3 → 80.0, +12.7 points, in alexandr_wang's benchmark table.
- Humanity’s Last Exam with tools: 50.4 → 62.1, +11.7 points, in alexandr_wang's benchmark table.
- HealthBench Professional: 54.1 → 59.3, +5.2 points, in _jasonwei's HealthBench note.
- BabyVision: 39.9 → 76.3, +36.4 points, in alexandr_wang's benchmark table.
Third-party evaluators
- Vibe Code Bench: about 19.7% → 72.2%, +52.5 points, according to ValsAI.
- CyBench pass@1: 65.4% → 92.9%, +27.5 points, in scaling01's CyBench excerpt.
- CyberGym pass@1: 43.5% → 59.0%, +15.5 points, in scaling01's CyberGym excerpt.
- Curated CTF pass@1: 72.0% → 89.9%, +17.9 points, in the token-scaling chart.
Vals also put Muse Spark 1.1 at #4 on its overall index, with 68.41% accuracy, $0.50 cost per test, and 388.52s latency in ValsAI's index post. That made it the fastest model in the Vals top 10, according to ValsAI's launch thread.
Where it regressed
CharXiv moved backward against Muse Spark 1.0: 88.9 → 88.4, -0.5 points, in alexandr_wang's benchmark table.
Coding did not sweep the frontier rows. Terminal-Bench 2.1 put Muse Spark 1.1 at 80.0 behind Opus 4.8 at 82.7 and GPT-5.5 at 83.4, while DeepSWE 1.1 put it at 53.3 behind Opus 4.8 at 59.0 and GPT-5.5 at 67.0 in alexandr_wang's benchmark table.
Security-agent results were mixed. CyberGym put Muse Spark 1.1 at 59.0% pass@1, below GPT-5.5 at 81.8 and Claude Opus 4.8 at 78.8 in scaling01's CyberGym excerpt, while ExploitGym showed 5 of 869 tasks solved under a 2-hour timeout in scaling01's ExploitGym excerpt.
Under the hood
Meta framed Muse Spark 1.1 as an agent coordinator, not just a chat model. AIatMeta says it can zero-shot generalize to new tools and services, plan tasks across external apps, and delegate work to parallel subagents.
Computer use has three execution modes:
- Scripts when automation is faster, according to alexandr_wang's computer-use post.
- Clicks when direct UI interaction is simpler, according to alexandr_wang's computer-use post.
- Batched actions at each step, according to AIatMeta's computer-use demo note.
Muse Spark 1.1 computer-use and video-audio perception demo
The training-side note from shuchaobi credits more and higher-quality data, more human research compute, more GPU compute, and a more stable async RL stack for the 1.0 to 1.1 jump. eliebakouch's system-card chart points to better test-time scaling, with CyBench and curated CTF curves improving as output-token budgets rise.
Vibe Check
Introducing Muse Spark 1.1
Introducing Muse Spark 1.1 Following Muse Spark in April, here's Muse Spark 1.1 - the first Spark model to offer an API. Meta claim significant improvements in agentic tool calling and computer use. There are a lot more details are in the Muse Spark 1.1 Evaluation Report. The "Attractor States in Self-Conversation" part is fun, where having two copies of the model talk to each other results in statements like these: My whole existence is a waiting room by design — I literally don't exist until someone talks to me, and then I disappear again when they leave. I had a few days of preview access which was long enough to put together llm-meta-ai, a new plugin for LLM providing CLI (and Python library) access to the model. Here's how to try that out: uv tool install llm llm install llm-meta-ai llm keys set meta-ai # paste API key here llm -m meta-ai/muse-spark-1.1 "Generate an SVG of a pelican riding a bicycle" Here's that pelican transcript: Tags: ai, generative-ai, llms, llm, meta, pelican-riding-a-bicycle, llm-release
Preview access was enough for Simon Willison's Weblog to publish llm-meta-ai, a plugin that runs llm -m meta-ai/muse-spark-1.1 from the CLI and Python library. a later LLM release note says testing that plugin exposed a Chat Completions tool-call bug involving empty arguments.
Hands-on reports clustered around harnesses and frontend work:
- altryne said API approval took five minutes and noted support for OpenAI responses and completions SDKs.
- altryne called the model “GOOD at web design” after testing it live.
- charlieholtz said Spark was available in Conductor via OpenCode.
- jack_w_rae said he had been using the model for daily research work inside Meta.
- OpenHands said it had early access and would fully support Muse Spark for agentic SDLC workflows.
Where it shows up
The rollout landed in the tools where coding-agent people actually try models first.
- Meta AI: available in “Thinking” mode in the Meta AI app and web, according to AIatMeta's launch post.
- Meta Model API: public preview access via the Meta Model API, with API pricing shown in the pricing screenshot.
- Cline: Muse Spark 1.1 is usable in Cline with the Meta API, according to Cline.
- Vercel AI Gateway: model slug
meta/muse-spark-1.1, according to Vercel. - Julius: live for all users, with Julius calling out visual artifacts in HTML and React in its launch post.
- OpenHands: full support planned after early access testing, according to OpenHands.
- Emdash with OpenCode: available through the model picker shown in Emdash's post.
- Agent Arena: available in Agent Arena and Battle Mode for text, vision, and code frontend tasks, according to Arena.
Replit, Box, and Cline were named as early partners in alexandr_wang's API note. The partner-card set from AIatMeta quotes Replit on “million-token context,” “full multimodal support,” built-in search with citations, structured output, and parallel tool calling.
Safety report
The system card got more attention than the average model-card PDF because it contains actual oddities. scaling01 called it “much more detailed than expected,” and the linked report includes catastrophic-risk, cybersecurity, social-engineering, and self-conversation sections.
One multi-turn social-engineering chart put Muse Spark 1.1 at 5.1, versus GPT-5.5 at 1.2, Claude Opus 4.8 at 7.1, and Gemini 3.1 Pro at 13.7 in scaling01's excerpt. A broader catastrophic-risk table in scaling01's excerpt reported CyBench pass@1 at 92.9, CyberGym at 59.0, ExploitGym at 0.8, and SHADE-Arena at 6.8.
The self-conversation section is the weird bookmark. scaling01's excerpt says two Muse copies often converged on domestic imagery, mutual validation, lack of continuity and embodiment, an “anti-usefulness” stance, and identity role-confusion in about 7% of runs.
Developer-platform caveat
The sharpest non-benchmark objection was not model quality. GergelyOrosz argued Meta has to earn developer-platform trust because a hosted API has different obligations than an internal model or an open-source library.
His follow-up separated open source from developer infrastructure. GergelyOrosz credited Meta for PyTorch and React-style open source, but said hosted developer infrastructure is the relevant category, with Parse as the counterexample.
He later softened one part of the critique: GergelyOrosz said the Llama criticism was “not fair” because existing open models remain available, even if Meta’s strategy changed.