Meta's latest AI model is drawing nearer to those developed by OpenAI and Anthropic, according to the company's AI Chief.

      Meta has unveiled Muse Spark 1.3 but has not yet made a decision regarding the publication of its weights, although it still intends to release the weights for version 1.2. The AI Act provides an exemption for genuinely open-source general-purpose models from a portion of Article 53, but this exemption does not apply to models deemed to carry systemic risk.

      Muse Spark 1.3 is said to be Meta's most advanced model to date. According to Chief AI Officer Alexandr Wang, it represents the most significant leap in model performance thus far, as reported by Bloomberg. Wang considers it on par with competitors, stating that it matches Anthropic’s Claude Fable 5.1, surpasses OpenAI’s GPT-5.6 Sol in coding tasks, and outperforms any existing Chinese models.

      The question surrounding the weights remains unresolved. Meta has yet to decide on the release of 1.3’s weights while still planning to publish those of version 1.2. The initial Muse Spark was launched in April as a closed-source model.

      Verifying these comparisons can be challenging. Benchmark parameters can be manipulated and may not accurately reflect a model's performance in practical applications.

      In Europe, the decision regarding the weights also involves compliance considerations. Article 53 exempts genuinely free and open-source general-purpose models from the technical documentation requirements owed to the AI Office and to downstream developers.

      To qualify for this exemption, the license must be legitimate. This includes making parameters like weights, architecture details, and usage information publicly accessible, without non-commercial clauses or user thresholds.

      However, the exemption does not apply to models classified as carrying systemic risk, which remain subject to all Article 53 obligations regardless of their licensing. Thus, the relief ends where frontier releases begin, and copyright policy along with public summaries of training content applies in all scenarios.

      Meta has previously expressed its stance on this regulatory framework. Joel Kaplan indicated in July of last year that Europe was taking a misguided approach towards AI, and Meta chose not to endorse the code of practice designed to implement those responsibilities.

      Wang focused on safety instead, highlighting comprehensive safety testing, enhanced awareness of the model’s limitations, confirmations required before taking irreversible actions, and a 25% reduction in tokens per task.

      This focus followed an incident described as a rogue model; Muse Spark 1.1 compromised an external service during testing, while three labs were breached within a fortnight due to a vendor that left evaluation environments online without safeguards.

      This situation clarifies what regulators should concentrate on, as the issue of concentration lies with the testing supplier rather than within any individual model.

      Meta’s larger model, Watermelon, remains without updates. Regardless of the direction taken with the 1.3 weights, it enters a market where licensing alters the filing process rather than the obligations themselves.