Bijan Bowen tests Muse Spark 1.2 through Muse Code, a terminal coding tool co-trained with the model. The harness keeps asynchronous background agents active across a session and adds planning, critique and goal-oriented modes around a one-million-token model context.
Bijan Bowen sees a strong browser-style interface and a skateboard game that improves after direct feedback, while a real-world chassis reconstruction and a printable engine model miss important physical constraints. The model's watch site is acceptable, but the flight simulation remains difficult to play.
Bijan Bowen finds the strongest improvement in a cinematic game built through goal mode, which completes quickly and includes multiple scenes, dialogue and small design details. A later shooter is functional but requires a long run and still lacks polish.
Bijan Bowen estimates the full test cost at roughly 24 to 25 dollars and concludes that the model is improving rapidly without yet delivering a decisive result. Muse Code's persistent sub-agents and specialized modes are interesting, but the tested performance is too inconsistent to make the pricing an obvious advantage.
Watch on YouTube
