Meta 发布 Muse Spark 1.3，称相比 1.2 工具调用减少 20%、生成 token 减少 25%，长编码任务可用更少的模型动作和输出来完成同一交付物。该版本在 MRCR 256K–512K 上得分 98.5，超过 GPT-5.6 Sol；在 JobBench（64.9 对 65.7）、OSWorld（66.9 对 68.3）和 AutomationBench（49.4 对 50.3）上接近 Opus 5。转发的官方公告称该版本当天在 Muse Code 和 API 上线，并预告开放权重版本即将发布。
对比表格列出 Muse Spark 1.3 与 GPT-5.6 Sol、Opus 5 在长上下文和智能体基准上的差距，可看出其性能定位。
Meta released Muse Spark 1.3. Reports 20% fewer tool calls and 25% fewer tokens than 1.2.
So long coding jobs should require fewer model actions and less generated text to reach the same deliverable.
- Long context is the standout: Muse Spark 1.3 scores 98.5 on MRCR 256K–512K. Beats GPT-5.6 Sol
- it gets very close to Opus 5 on JobBench (64.9 vs 65.7), OSWorld (66.9 vs 68.3) and AutomationBench (49.4 vs 50.3)
Muse Spark 1.3 is rolling out today with frontier performance almost too cheap to meter. This is the biggest jump we've made so far on coding and agentic work. Try it in Muse Code and our API. Next up 🍉 and Muse Spark open weights releases coming soon. https://t.co/XQQEDEJGD7在 X 查看被引用的帖子
来源：@rohanpaul_ai · x.com