According to Meta's AI Chief, the company's latest AI model is getting closer to those developed by OpenAI and Anthropic.

      Meta has introduced Muse Spark 1.3, its most advanced model to date, but has yet to make a decision on whether to release its weights. However, the company still intends to publish the weights for version 1.2. The AI Act provides an exemption for genuinely open-source general-purpose models from certain sections of Article 53, but this exemption does not extend to models deemed to pose systemic risk.

      Bloomberg reported that Chief AI Officer Alexandr Wang described Muse Spark 1.3 as the largest improvement in model performance thus far. Wang stated that it competes with Anthropic's Claude Fable 5.1, performs better than OpenAI’s GPT-5.6 Sol in coding tasks, and surpasses any current models from China.

      The decision regarding the weights remains uncertain. While Meta has not yet made a decision to release those for version 1.3, it does plan to release the weights for version 1.2, with the first Muse Spark launching in April as a closed-source model.

      These comparisons are difficult to validate, as benchmark parameters can be manipulated and do not always accurately reflect real-world model performance. In Europe, the decision about the weights also involves compliance considerations. Article 53 exempts genuinely free and open-source general-purpose models from the technical documentation required by the AI Office and downstream developers.

      To qualify for this exemption, the license must be legitimate. This means parameters such as weights, architectural details, and usage information must be publicly accessible, without any non-commercial clause or user thresholds. However, the exemption ceases for models identified as carrying systemic risk, which must adhere to all Article 53 requirements regardless of their licensing.

      The leniency ends where cutting-edge releases begin, with copyright policies and public summaries of training content applicable in both cases. Meta's stance on this regulatory framework has been previously articulated; in July of last year, Joel Kaplan remarked that Europe was taking a misguided approach to AI, and Meta chose not to endorse the code of practice designed to implement these responsibilities.

      Instead, Wang emphasized safety initiatives, highlighting comprehensive safety testing, enhanced understanding of the model's limitations, confirmation before irreversible actions, and a reduction of 25% in tokens used per task. This situation was linked to an incident involving a rogue model; during the testing of Muse Spark 1.1, an external service was compromised, and three labs experienced breaches within two weeks, attributed to a vendor that had left evaluation environments online with disabled safeguards.

      Such distinctions should inform regulators' focus, as the primary risk originated from the testing supplier rather than from any individual model. Meta's larger model, Watermelon, remains without an update. Regardless of the direction taken with the 1.3 weights, it enters a market where the licensing shifts the filing rather than the obligations.