July 9, 2026 wasn't a quiet day for Meta. On the same day OpenAI made GPT-5.6 generally available, Meta raised its own stakes in the coding and agentic-AI market by releasing Muse Spark 1.1.
A New Vision Led by Alexandr Wang
Muse Spark 1.1 was built under Meta Superintelligence Labs (MSL), led by Alexandr Wang. As a direct follow-up to the original Muse Spark introduced in April 2026, it's a concrete signal of how much weight Meta's AI strategy now puts on coding and on agents that "don't just think, they act."
In Meta's own words: "Meta AI doesn't just think, it acts." That philosophy shows up directly in Muse Spark 1.1's technical spec sheet.
Standout Features
- 1-million-token context window — capable of processing large codebases or long document sets in a single pass.
- Multi-agent orchestration — the model can break complex projects across multiple sub-agents, optimized to reduce end-to-end latency.
- Major gains in tool and computer use — a clear jump over the original Muse Spark in tool-use and computer-use capability.
- Strong performance on personal agentic tasks requiring planning and coordination — it stands out on tasks that require orchestrating across external apps and services.
Meta announced that Muse Spark 1.1 beats Google's latest Gemini release on some coding and reasoning benchmarks — a claim that lands with extra weight given Google was simultaneously delaying its own flagship, Gemini 3.5 Pro, over coding shortfalls.
A Closer Look at the Benchmark Table
Digging a bit deeper into that claim: Muse Spark 1.1 scored 77.4% on SWE-bench Verified — behind Claude Opus 4.6's 80.8% and Gemini 3.1 Pro's 80.6%, though not by a huge margin. On SWE-bench Pro, Meta's own reported score is 61.5%, notably behind the leading models on that particular benchmark.
The real differentiator shows up not in pure coding tests, but in tool use and agent orchestration: Muse Spark 1.1 scored 88.1 on the MCP Atlas test, ahead of Claude Opus 4.8's 82.2, putting it in the lead on agentic, tool-orchestration tasks. In other words, Meta's "we beat Gemini" claim rests less on raw coding ability and more on the model's skill at coordinating tools and external services — which lines up neatly with the company's own "doesn't just think, it acts" framing.
Muse Spark 1.1 was just the opening move in Meta's rapid iteration pace throughout summer 2026. A few weeks later, on August 5, the company shipped Muse Code, a terminal-based coding agent, alongside Muse Spark 1.2; on August 10, it announced Muse Glimmer, a smaller, open-weight model.
That cadence reflects an aggressive strategy: Meta doesn't want to fall behind Anthropic and OpenAI in the coding and agentic-AI market — and Muse Spark 1.1 was the first hard evidence of that strategy in motion.
