Meta is preparing an open-weights release for Muse Spark, according to Mark Zuckerberg. Techmeme’s summary of Zuckerberg’s @finkd post says open weights for Meta’s Watermelon model and Muse Spark are “coming soon,” while The Register separately reports that Zuckerberg promised an open-weights version of Muse Spark “soon” in a Wednesday X post. The part that is already live is Muse Spark 1.3. Techmeme’s summary says the model is rolling out, and The Register reports that Spark 1.3 is now available through Meta’s API service and the Muse Code command-line interface. The Register describes Spark 1.3 as a refinement of the Muse Spark model Meta released in April. Its report says Meta has continued to update the model as it tries to keep pressure on rivals and justify heavy AI capital spending to investors. The most concrete changes are aimed at agentic and coding workflows. According to The Register, Meta says Spark 1.3 should perform better on long-horizon tasks and waste fewer tokens by repeatedly pursuing unproductive paths. The model is also described as more likely to ask clarifying questions, request human help when it gets stuck, and confirm before taking consequential actions. That matters operationally because multi-step agent runs can become expensive when a model loops, guesses, or presses ahead without enough context. The Register says Meta frames the changes as reducing the number of turns and tokens needed to finish a task, which would make the model cheaper to use in practice if the behavior holds up outside Meta’s own examples. On performance, The Register says Meta claims gains over Muse 1.2 across a range of benchmarks. It also cites Artificial Analysis results showing a 4-point increase in overall intelligence, putting Spark 1.3 in a dead heat with GPT 5.6 Sol, Claude Opus 5, and Grok 4.6 High. Those benchmark details are single-sourced here to The Register’s report. The Register also reports discounted pricing for Meta’s contributor tier: as low as $0.002 for cached input, $0.10 for input, and $0.20 for output per million tokens. Its caveat is that this tier comes with lower rate limits and a tradeoff around Meta using prompts for training data. The Watermelon detail is thinner. Techmeme’s item says Zuckerberg referred to Watermelon open weights as coming soon, but the provided Register report does not discuss Watermelon. For now, the better-supported story is Muse Spark 1.3’s availability and Zuckerberg’s signal that a Muse Spark open-weights release is on the way. Who benefits: Developers using Meta’s API service or Muse Code CLI get immediate access to Spark 1.3. Teams experimenting with AI agents and code assistants may benefit if the reported reductions in unnecessary turns and token use hold up in their workloads. Who's exposed: Rival frontier-model providers face another low-cost, performance-focused release from Meta. Users considering Meta’s discounted contributor tier should weigh the reported rate-limit and prompt-use tradeoffs described by The Register.