Thesis
Muse Spark Open treats long-running coding-agent work as a model architecture and serving problem, not merely a matter of extending context length.
Observed facts
Baseten says Muse Spark Open has 816B total parameters, 42.7B active at inference, a 1M-token context window, and accepts text, image, video, and audio input.
1Baseten describes 46 local-attention layers and 16 global-attention layers, and estimates KV-cache use at 128K context falls from 31 GiB to 8.4 GiB per sequence.
1Baseten reports its NVFP4 checkpoint is about 540 GB versus roughly 1.6 TB and fits on one 4×GB300 node.
1
Mechanism
Baseten describes a compressed MoE block intended to reduce parameter footprint and inter-GPU traffic, alongside 46 local-attention layers and 16 global-attention layers. At 128K context, the post estimates KV-cache use falls from 31 GiB to 8.4 GiB per sequence.
Why now
The release makes the weights available and links model design to the repeated inference and memory demands of agent loops.
Implications
The first implication is that teams evaluating open-weight models for coding agents can assess not only task quality but also whether the model’s context and serving footprint fit their workloads.
Evidence boundary
The source is Baseten’s announcement; benchmark comparisons are attributed to Meta and use lab-specific agents rather than one shared harness, so they do not establish a controlled ranking across models.
Watch conditions
What to watch: whether independent teams can reproduce the stated throughput and memory advantages on comparable agent workloads.