Meta releases Muse Spark 1.3, tops DeepSWE at 75.4% and opens paid API access

Muse Spark 1.3 is the clearest signal yet of Meta's strategic overhaul: after concluding that its open-source Llama models weren't keeping pace with frontier rivals, Zuckerberg pushed the team toward a closed, proprietary, directly monetized product, with paid API access opening September 3. It's the fourth release in roughly five months — a cadence meant to close the gap fast. Alexandr Wang framed it as bringing significant advances in coding and agentic capabilities, and a widely-shared benchmark claim put Muse Spark 1.3 first on DeepSWE at 75.4%, ahead of GPT-5.6 Sol and Fable 5.

The competitive positioning is explicit: Meta wants to be evaluated against Anthropic and OpenAI's frontier tiers, not open-weight peers. Meta also promised a future model codenamed Watermelon and said open-weight Muse Spark releases are still coming, a hedge to avoid fully alienating the Llama community that built its ecosystem. Separately it released Muse Voice Transcribe for real-time transcription.

The risk is community backlash and enterprise trust. Developers who benefited from free Llama models are questioning the long-term impact of the proprietary pivot, and Meta is simultaneously trying to win back enterprises that drifted away when Llama fell behind. There's irony in the strategy: Meta has quietly become one of Microsoft's largest Azure AI customers, spending heavily on rival models to evaluate its own — underscoring how far it had fallen and how much it's spending to catch up. Whether the DeepSWE-topping benchmark translates into real production adoption, and whether the promised open-weight releases materialize, will determine if the pivot restores Meta's standing or just monetizes a shrinking mindshare.