模型與代理
Muse Spark 1.3 Reduces Agent Tool Calls, While Its Highest Reasoning Mode Still Awaits Safety Testing
Meta has updated Muse Spark with a focus on long-horizon agents, software development, and task tracking in mixed-topic conversations. Internal comparisons show reductions in both tool calls and token usage. The model is now available in Muse Code and via the API, but its highest reasoning mode, open weights, and fully reproducible evaluations have yet to arrive.
Meta launched Muse Spark 1.3 on September 2, making it available through Muse Code and the Meta Model API. Rather than simply increasing the parameter count or context window, this update retrains the model to perform long-horizon tasks across multiple agent harnesses. It can build working context from conflicting sources, fill gaps in plans, identify the current instruction within a single conversation containing multiple interleaved tasks, and is more likely to ask users for confirmation when requirements are unclear, it encounters obstacles, or it is about to perform an irreversible action.
For software development work, Meta says that in internal comparisons by engineers, Spark 1.3 used roughly 20% fewer tool calls and 25% fewer tokens than Spark 1.2. It was also trained more extensively on long-running coding tasks. If these differences can be reproduced using the same harness, model temperature, and stopping conditions, they could matter more to agent costs and cumulative failure rates than gains on any individual benchmark. Every eliminated tool round trip also removes one opportunity for a permissions, network, or parsing error.
However, the official evaluation combines Meta’s own test runs, public leaderboards, and competitors’ self-reported best scores. Third-party models were also tuned only on a “best-effort” basis, so the comparisons are not fully symmetrical. The OSWorld versions used for Spark 1.2 and the new model are inconsistent, while the internal IF Index has no publicly available fixed test set. Meta claims improved resistance to prompt injection and better judgment around irreversible actions, but it has not published independently verified attack success rates. More notably, the existing reasoning modes launched first, while `max` mode still requires additional safety testing. The promised open weights have not yet been released either. Engineering teams should next monitor real-world API pricing, latency, and tool error rates, as well as whether the highest reasoning mode is accompanied by an updated safety report when it launches.