Baseten serves Meta's Muse Spark Open at launch

As we reported earlier, Meta has released the weights for Muse Spark Open, its open-weight model for coding agents. Inference provider Baseten now says the model is available on its Model APIs and Dedicated Inference products. Baseten serves an NVFP4 checkpoint, which it says cuts the weight footprint by roughly two-thirds. That leaves more GPU memory for KV cache and for concurrent agent sessions.

The model is a mixture-of-experts (MoE) design with 816B total parameters, 42.7B of which are active during inference. It has a 1M-token context window and accepts text, image, video and audio input, with text output. Meta co-trained it with its Muse Code agent. Baseten says it also works with third-party agents, including OpenCode, Pi and Deep Agents Code. Baseten calls it the most capable US open-weights model to date.

Baseten's post also explains the model's architecture:

• Compressed MoE block: a shared pooler shrinks each token's input before it reaches the experts, while the router still reads the full-width vector. This halves each expert's parameters, saving roughly 800B parameters compared with an uncompressed design, and reduces traffic between GPUs.
• Hybrid attention: 46 of the model's 62 layers look only at the most recent 2,048 tokens. The other 16 layers attend to the full sequence. Baseten says this cuts KV-cache memory by about 3.7x at 128K context.

Baseten's summary of Meta's reported benchmarks says the model comes close to flagship closed models on routine agent loops. On large multi-file changes it is expected to need more guidance or retries than Claude Opus 5 or GPT-5.6 Terra. On write-up tasks, Claude Opus 5 is the only model in Meta's comparison that scores higher. Each model was tested with its own lab's coding agent, not a shared harness. The post does not give the raw scores.