Muse Spark Open added to BenchGecko · Meta 816B open-weight MoE with 42B active and 1M context
Meta released open weights for Muse Spark Open on 2026-10-15. BenchGecko lists it at $1.25 input and $4.25 output per M tokens with a 1,048,576 token context window.
What launched
Meta released Muse Spark Open on 2026-10-15 under the Muse Model License. It is the open-weight member of the Muse family, which started with Muse Spark on 2026-04-08. Weights are on Hugging Face as BF16 and NVFP4 checkpoints, and seven API platforms serve it at launch: Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI.
BenchGecko added the model today and corrected the older Muse Spark row at the same time: provider set to Meta and release date set to 2026-04-08.
Specs
From Meta's model card:
- Architecture · Mixture-of-Experts Transformer with a Perception Encoder, 816B total parameters, 42B active per token
- Experts · 256 per MoE layer, 8 active per token
- Depth · 62 layers, 128 query heads and 16 KV heads
- Context · 1M tokens, stored as 1,048,576 on BenchGecko to match provider listings
- Inputs · text, images, video and audio
- Output · text
- Vocabulary · 202,048 tokens
- Reasoning effort · low, medium, high or xhigh, with xhigh as the default
Self-hosting
The model card asks for at least 700GB of GPU memory for the NVFP4 checkpoint and 2TB for BF16. Meta's developer docs list weight footprints of about 537 GB for NVFP4 and 1.65 TB for BF16, before KV cache. Meta also tuned it to run on two Mac Studios with M5 Ultra through MLX. Supported runtimes include SGLang, vLLM, TensorRT-LLM, Transformers and llama.cpp.
Licence
The Muse Model License allows commercial use, fine-tuning and distillation. It requires a visible "Built with Muse Spark Open" notice, and Model as a Service providers above $20 million in yearly revenue need a separate agreement with Meta.
Vendor-reported benchmarks
Meta reports these results at xhigh reasoning effort. BenchGecko records them as vendor claims and has no Gecko Tests for this model yet.
- Artificial Analysis Intelligence Index · 39
- MCP-Atlas · 89.2
- OSWorld Verified · 80.4
- Terminal-Bench 2.1, Terminus 2 agent · 76.4
- Toolathlon Verified · 72.8
- GPQA Diamond · 89.0
- HLE with tools · 61.0
- HLE-Diamond, no tools · 26.1
- SciCode · 58.5
- JobBench · 61.8
- GDPVal-AA v2 · 1401
In the same table, Mistral Large 4 is listed at 38 on the Intelligence Index and 54.2 on SciCode. GLM 5.3 at max effort is listed at 45 and Kimi K3 at max effort at 44 on the Intelligence Index.
Pricing on BenchGecko
Meta's card says API access runs through third-party inference providers. The five paid OpenRouter endpoints (Together, Fireworks, Baseten, Modal and Nebius) list the same price, so BenchGecko records $1.25 input, $0.15 cached input and $4.25 output per M tokens. Modal serves 1,000,000 tokens of context and the others 1,048,576. For comparison, Muse Wonder, Meta's closed frontier model, is listed at $4 input and $20 output per M tokens.
Sources
- Meta announcement
- Model card on Hugging Face
- Meta developer docs
- OpenRouter endpoints
- Muse Spark launch post
- BenchGecko model page: /model/muse-spark-open
Source: BenchGecko
Written automatically from BenchGecko's tracked data for this period; every number comes from the changelog, model, mindshare and pricing datasets. Checked against the source data before publishing.