Meta released the weights for Muse Spark Open on October 15, 2026, and Baseten made the 816B-parameter open-weight omnimodal model available through Model APIs and Dedicated Inference. Co-trained with Muse Code, it is designed for coding-agent workflows and has 42.7B active parameters during inference, a 1M-token context window, and text, image, video, and audio input with text output. Baseten says it is the most capable US open-weights model to date. The model can work with Muse Code and third-party agents including OpenCode, Pi, and Deep Agents Code.

Meta-reported benchmark scores place Muse Spark Open among models from other frontier labs on agentic coding and knowledge work. The comparisons used each model with its own lab’s coding agent: Muse Spark with Muse Code, Claude Opus 5 with Claude Code, and GPT-5.6 Terra with Codex, rather than a shared test setup. The comparison includes each lab’s flagship model but excludes higher-tier Claude Fable 5.1 and GPT-6 Astra. Baseten says Muse Spark Open performs close to flagship closed models in routine command-running and debugging loops, but needs more guidance or retries on large, multi-file changes than Claude Opus 5 or GPT-5.6 Terra. It holds up well against Grok 4.5 and Gemini 3.6 Flash, and leads every model in the chart except Opus 5 on written work summaries.

To reduce compute and memory demands in long agent sessions, the model combines a compressed Mixture-of-Experts (MoE) block, which routes tokens to selected expert networks, with hybrid global and local attention. A shared pooler compresses each input vector before it reaches the experts, and a shared unpooler restores their combined output; the router still reads the full-width vector. This halves each expert’s parameters and reduces total model size by roughly 800B parameters versus an uncompressed design, freeing GPU memory for larger KV caches and concurrent sequences while reducing inter-GPU traffic.

Of its 62 layers, 46 use local attention with a 2048-token sliding window and rotary position embeddings (RoPE), while 16 global-attention layers process the full sequence without positional encoding (NoPE). At 128K tokens of context, this schedule cuts KV-cache memory per sequence from 31 GiB to 8.4 GiB, about 3.7x less than a model using only global-attention layers; Baseten says savings grow with context length. Earlier tokens outside a local layer’s window can be overwritten there, while global layers retain full-context information in their KV cache.

Baseten serves an NVFP4 checkpoint, storing expert weights in 4-bit and using bf16 for attention weights and experts in the first and last layers. It reduces model weights from roughly 1.6 TB to 540 GB, about two-thirds less, and fits the model on one 4×GB300 node without a multi-node interconnect. Baseten says the freed memory allows more concurrent agent sessions per node and higher throughput. The OpenAI-compatible Model API supports tool calling, reasoning output, and the full 1M-token context window; teams can also use Dedicated Inference with reserved capacity.

Meta released the weights for Muse Spark Open on October 15, 2026, and Baseten made the 816B-parameter open-weight omnimodal model available through Model APIs and Dedicated Inference. Co-trained with Muse Code, it is designed for coding-agent workflows and has 42.7B active parameters during inference, a 1M-token context window, and text, image, video, and audio input with text output. Baseten says it is the most capable US open-weights model to date. The model can work with Muse Code and third-party agents including OpenCode, Pi, and Deep Agents Code.

Meta-reported benchmark scores place Muse Spark Open among models from other frontier labs on agentic coding and knowledge work. The comparisons used each model with its own lab’s coding agent: Muse Spark with Muse Code, Claude Opus 5 with Claude Code, and GPT-5.6 Terra with Codex, rather than a shared test setup. The comparison includes each lab’s flagship model but excludes higher-tier Claude Fable 5.1 and GPT-6 Astra. Baseten says Muse Spark Open performs close to flagship closed models in routine command-running and debugging loops, but needs more guidance or retries on large, multi-file changes than Claude Opus 5 or GPT-5.6 Terra. It holds up well against Grok 4.5 and Gemini 3.6 Flash, and leads every model in the chart except Opus 5 on written work summaries.

To reduce compute and memory demands in long agent sessions, the model combines a compressed Mixture-of-Experts (MoE) block, which routes tokens to selected expert networks, with hybrid global and local attention. A shared pooler compresses each input vector before it reaches the experts, and a shared unpooler restores their combined output; the router still reads the full-width vector. This halves each expert’s parameters and reduces total model size by roughly 800B parameters versus an uncompressed design, freeing GPU memory for larger KV caches and concurrent sequences while reducing inter-GPU traffic.

Of its 62 layers, 46 use local attention with a 2048-token sliding window and rotary position embeddings (RoPE), while 16 global-attention layers process the full sequence without positional encoding (NoPE). At 128K tokens of context, this schedule cuts KV-cache memory per sequence from 31 GiB to 8.4 GiB, about 3.7x less than a model using only global-attention layers; Baseten says savings grow with context length. Earlier tokens outside a local layer’s window can be overwritten there, while global layers retain full-context information in their KV cache.

Baseten serves an NVFP4 checkpoint, storing expert weights in 4-bit and using bf16 for attention weights and experts in the first and last layers. It reduces model weights from roughly 1.6 TB to 540 GB, about two-thirds less, and fits the model on one 4×GB300 node without a multi-node interconnect. Baseten says the freed memory allows more concurrent agent sessions per node and higher throughput. The OpenAI-compatible Model API supports tool calling, reasoning output, and the full 1M-token context window; teams can also use Dedicated Inference with reserved capacity.