Muse Spark Open
research.meta.ai
October 15, 2026
3 min read
74/100
Summary
Meta has released the weights for Muse Spark Open, an open-weight, natively multimodal AI model designed for coding, agentic workflows, complex software engineering, and multi-step reasoning. The model has 816 billion total parameters, with 42 billion active parameters, and supports a context window of up to 1 million tokens. Meta calls it the most capable US open-weight model to date. Muse Spark Open can run locally on two Apple M5 Ultra Mac Studio systems using Apple’s MLX framework. Each M5 Ultra can provide up to 512 GB of unified memory and 1.2 TB/s of memory bandwidth, allowing the two systems to hold the full model. Meta tested on-device workflows with Muse Code, including making bug fixes and adding features to the Ice Cubes Mastodon client, as well as reviewing and vertically cropping video in DaVinci Resolve. The weights are available through Amazon SageMaker AI, Google Cloud Model Garden, Hugging Face, and Microsoft Foundry Models, while hosted inference is offered through Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, and Together AI. The model supports engines including SGLang, vLLM, TensorRT-LLM, Transformers, MLX, llama.cpp, and Unsloth. Meta says its SGLang kernels deliver 1.03-3.95 times kernel-level speedups on NVIDIA and AMD GPUs. The model is released under the Muse Model License with limited commercial terms.
Key Takeaways
- Meta released Muse Spark Open, an 816-billion-parameter multimodal open-weight model with 42 billion active parameters and a 1-million-token context window.
- Two Apple M5 Ultra Mac Studio systems can run Muse Spark Open locally through the MLX framework, using up to 1 TB of combined unified memory.
- Muse Spark Open is available for download through Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden, and Microsoft Foundry Models, with hosted access from seven inference providers.
- Meta says Muse Spark Open received additional safety post-training and met its Superintelligence Scaling Framework thresholds for release.
- The weights use the Muse Model License, whose limited commercial terms allow many developers to use the model for free.
What the discussion said
Commenters treated Muse Spark Open less as a breakthrough than as a test of whether Meta can still matter in open-weight models. The benchmark-heavy view was blunt: its reported high-reasoning score lands around DeepSeek V4.1 Flash and Mimo 2.5 Flash, but behind Qwen 3.8 Flash Next and GLM 5.3 Flash. That gap stings because those competitors are said to fit on a single B300 and, according to one cost comparison, can be far cheaper to serve. Several readers also argued that max-reasoning leaderboards hide the low- and medium-effort settings that determine real production throughput, latency, and cost. There was real credit for the release. Meta documented deployment across major local runners instead of dumping weights on the community, and a hands-on image-to-HTML trial found it unusually strong among open models even if it missed details. Some see a capable US/EU-controlled model as valuable for customers needing sovereignty, and Meta’s consumer reach could make an AI product broadly adopted even without technical leadership. But the open label drew sustained resistance. The license forces sufficiently large model-serving businesses into a separate agreement, which readers say constrains provider competition and disqualifies it from genuinely open-source use. Claims of democratized local AI also rang hollow to commenters when the headline setup requires two extremely expensive Mac Studios. The thread’s prevailing verdict: useful release, but neither performance nor accessibility justifies the marketing sweepstakes yet.
Where opinion split
The sharpest fight was whether Meta can build a durable advantage around Muse despite trailing rival open-weight models. Optimists argue that price eventually dominates once models clear a capability threshold, and Meta can turn massive distribution, polished deployment, and ecosystem investment into a moat. Skeptics counter that benchmarked capability, inference cost, permissive licensing, and control of consumer platforms favor Chinese rivals, Google, Apple, and Microsoft—not a merely sticky Meta interface.
Community Sentiment
Positives
- Meta tested and documented deployment across runners such as Unsloth and MLX, making the weights materially easier to put into real local and hosted AI stacks.
- A hands-on image-to-HTML comparison found Spark Open captured the target design’s overall feel better than most open alternatives, even while missing fidelity details.
- A capable Western open-weight option has practical value for organizations that need model sovereignty or cannot deploy Chinese technology.
- Meta’s consumer-scale distribution could turn a merely competent model into a widely used AI product through rapid experimentation and integration.
Concerns
- Reported benchmark results place Spark Open behind Qwen 3.8 Flash Next and GLM 5.3 Flash, leaving Meta visibly behind the strongest open-weight competition.
- Comparable DeepSeek-level capability reportedly comes with much higher inference cost, a fatal trade-off when production buyers optimize for tokens, latency, and margin.
- High-reasoning benchmark scores are a poor proxy for production workloads, where low and medium effort settings decide whether an AI system is fast and economical enough to use.
- The revenue-triggered serving restriction weakens the open claim and may suppress provider competition that would otherwise push model access prices down.
- Calling local deployment democratized feels absurd when the promoted full-model setup needs two top-end Mac Studios and roughly a terabyte of unified memory.
