Meta Releases Muse Spark Open, an 816B Open-Weight AI Model

Meta has released open weights for Muse Spark, the large language model that debuted in April 2026 as the first product of its Meta Superintelligence Labs. The new version, called Muse Spark Open, went live on Hugging Face on October 15, 2026, giving developers free access to a model Meta calls "the most capable US open-weight model to date."

That claim puts Muse Spark Open in direct competition with China's leading open-weight releases — Alibaba's Qwen family, Moonshot AI's Kimi K3 and Z.ai's GLM models — which have dominated open-weight leaderboards through most of 2026. It's also Meta's clearest signal yet that it isn't abandoning open models entirely, even as its flagship consumer products move toward closed, proprietary releases like Muse Wonder.

What is Muse Spark Open?

Muse Spark Open is the open-weight release of Muse Spark, the model Meta Superintelligence Labs introduced in April 2026 as the first entry in its "Muse" series. Where the original Muse Spark launched as a closed model, available only through Meta's own apps and a private-preview API, Muse Spark Open publishes the underlying weights for anyone to download and run.

According to its Hugging Face model card, Muse Spark Open is a mixture-of-experts transformer built for "coding and agentic work": multi-step reasoning, end-to-end agentic task completion, tool use, multimodal reasoning, synthetic data generation and use as an LLM-as-a-judge evaluator. Meta explicitly recommends deploying it inside a system with its own guardrails rather than as a standalone model.

The backstory: Meta Superintelligence Labs and the original Muse Spark

Muse Spark first launched in a closed form on April 8, 2026, as the debut product of Meta Superintelligence Labs, the AI research group Meta formed after a company-wide reorganization of its AI efforts earlier that year under chief AI officer Alexandr Wang. At launch, Meta described Muse Spark as the successor to its Llama line and said it would power Meta AI, Meta's smart glasses, and AI features across Facebook, Instagram, WhatsApp and Messenger. The company said it worked with more than 1,000 physicians to improve the model's health-related reasoning, and it positioned a planned "Contemplating" mode — which runs several agent instances in parallel on harder problems — as a rival to the extended-reasoning modes offered by competing labs.

That original release was proprietary from day one, a shift Meta itself acknowledged at the time: for most of the preceding decade, Meta had marketed itself as the industry's standard-bearer for openly available models through the Llama family. Muse Spark Open, six months later, is Meta partially reversing that closed-first approach for at least one model in the family, while keeping its newer, more consumer-facing Muse Wonder models closed.

Specs: parameters, context window and architecture

That 1-million-token context window and native multimodal input (text, image, video and audio) put Muse Spark Open's raw specs ahead of most other open-weight releases in 2026, though raw parameter count alone doesn't translate directly into task performance.

Benchmark scores Meta has published

Meta's own model card lists scores run at the model's highest "xhigh" reasoning setting, including an Artificial Analysis Intelligence Index of 39, an MCP-Atlas score of 89.2, OSWorld Verified at 80.4, GPQA Diamond at 89.0 and Terminal-Bench 2.1 at 76.4. On safety evaluations, Meta reports a 99.2% jailbreak refusal rate on its internal cyber benchmark and a 0% misbehavior rate on its Agentic Misalignment test.

These are Meta's self-reported numbers, not independently audited third-party results, and Meta's own card notes they apply only to the specific released checkpoint under the exact conditions tested. Independent trackers that rank Muse Spark Open against Qwen, Kimi K3 and GLM place it in the upper tier of open-weight models above 150 billion parameters, but not at the very top of the Artificial Analysis Intelligence Index.

Where it ranks against Qwen, Kimi K3 and GLM

Meta frames Muse Spark Open as "the most capable US open-weight model to date," a claim aimed squarely at closing the gap with Chinese open-weight leaders. Independent trackers that log every open-weight release on the Artificial Analysis Intelligence Index — the composite benchmark score Meta itself cites on its model card — place Muse Spark Open's score of 39 roughly in the upper third of open-weight models above 150 billion parameters, but behind the current leaders from Alibaba's Qwen family, Moonshot AI's Kimi K3 and Z.ai's GLM 5.3 on that same composite score and on terminal/agentic coding benchmarks specifically.

Where Muse Spark Open does stand out is tool use and computer-operation tasks: its 89.2 on MCP-Atlas and 80.4 on OSWorld Verified are both near the top of the open-weight field, consistent with Meta's stated focus on building a model for "coding and agentic work" rather than chasing the single highest general-reasoning score. In short, Meta isn't claiming the outright top spot on every leaderboard — it's claiming the best US-made alternative to China's open-weight frontier, with tool-use and agentic benchmarks as its strongest evidence.

Safety testing and limitations Meta discloses

Meta's model card lists safety evaluations across four categories: content safety, agentic risk, privacy, and "preparedness" risks such as chemical, biological and cyber misuse, plus loss-of-control scenarios. Reported figures include a 99.2% refusal rate on an internal cyber-misuse benchmark, a 98.1% honesty score on the MASK benchmark, and a 0% misbehavior rate on Agentic Misalignment testing. Meta also ran adversarial "FORTRESS" jailbreak attempts against the model, reporting single-digit attack and override success rates.

Meta is explicit that these figures describe only the specific released checkpoint, under the exact conditions it tested, and that quantized community versions of the model (the compressed formats many developers will actually run) "may differ slightly in edge cases." The model card also states Muse Spark Open isn't intended for use by anyone under 18, and that very short or low-resolution video inputs may trigger refusals rather than analysis.

The license: open weights, not open source

Muse Spark Open ships under Meta's "Muse Model License," not a standard open-source license like Apache 2.0 or MIT. The license permits using, copying, modifying, fine-tuning, merging and redistributing the model, but attaches conditions: products built on it must display "Built with Muse Spark Open," and any company offering the model as a hosted API service that earns more than $20 million in gross revenue over any 12-month period must notify Meta and sign a separate commercial agreement. Internal use and non-API products are exempt from that threshold.

That structure mirrors the license Meta has used for its Llama models since 2023: free for almost all developers, but with strings attached for the largest commercial API providers. It means Muse Spark Open is "open-weight," in that anyone can download and inspect the weights, but it isn't open-source in the sense the Open Source Initiative uses the term.

Where to download and run it

Meta has published Muse Spark Open's weights on Hugging Face under the Meta Superintelligence Labs organization, with mirrors on Amazon SageMaker AI, Google Cloud Model Garden and Microsoft's Foundry Model Catalog. Supported inference frameworks include vLLM, SGLang, Hugging Face Transformers, TensorRT-LLM, Unsloth and llama.cpp.

Running the full model requires serious hardware: Meta's card lists roughly 700GB of VRAM for the compressed NVFP4 format or around 2TB for full-precision BF16 — well beyond a single consumer GPU. Meta has separately highlighted that it runs on two Apple Mac Studios with M5 Ultra chips and 512GB of unified memory each, using Apple's MLX framework, as a more attainable (if still expensive) local setup for developers who want to self-host it.

How it differs from Muse Wonder

Muse Spark Open shouldn't be confused with Muse Wonder and Muse Wonder Full, the closed, consumer-facing "frontier models for agents" that Meta Superintelligence Labs introduced on October 12, 2026. Muse Wonder powers Meta's own AI app, Instagram, WhatsApp, Facebook and Messenger features, and isn't available as downloadable weights. Muse Spark Open, by contrast, is the version developers can actually download, inspect and self-host — Meta's play for the open-weight developer and research community rather than its own consumer products.

Why Meta is doing this now

For years, Meta's Llama models were the default reference point for "open" AI in the West, while Chinese labs such as Alibaba, Moonshot AI and Z.ai (formerly Zhipu AI) spent 2025 and 2026 pulling ahead on open-weight benchmark leaderboards with frequent, well-documented releases. Muse Spark Open is Meta's attempt to reclaim that ground: a US-made, high-parameter, natively multimodal model released with a full model card, published benchmarks and safety evaluations, timed squarely against its Chinese open-weight rivals.

It also shows Meta running two AI tracks in parallel: closed, product-focused models like Muse Wonder for its own apps, and open-weight releases like Muse Spark Open aimed at outside developers, researchers and smaller companies that can't afford frontier-lab API pricing but want a capable base model to build on.

What developers get, in practice

Because Muse Spark Open's full-precision weights need roughly 2TB of memory to run, most individual developers won't be loading the raw checkpoint on their own hardware. In practice, access falls into three tiers: cloud rental through Google Cloud Model Garden, Amazon SageMaker AI or Microsoft Foundry (billed by compute time, with no separate model license fee below the $20 million API-revenue threshold); quantized community builds in formats like GGUF, AWQ and MLX that shrink memory requirements at some cost to accuracy, which third parties began uploading to Hugging Face within days of the official release; and the two-Mac-Studio local setup Meta itself demonstrated, aimed at researchers and well-funded labs rather than hobbyists.

That tiered reality matters for anyone deciding how to actually use Muse Spark Open: companies large enough to need an exact, unmodified copy of Meta's benchmarked checkpoint will likely rent cloud compute, while smaller teams experimenting with it will mostly run community-quantized versions whose real-world performance can diverge from Meta's published numbers.

What to watch next

Meta has signaled more Muse-family releases are coming, including further iterations tied to Muse Spark's internal version numbers. Independent benchmark trackers will keep testing Muse Spark Open against Qwen, Kimi K3 and GLM over the coming weeks as more developers get hands-on access, which should give a clearer, less Meta-controlled picture of where it actually lands. Anyone evaluating it for a commercial product should read the Muse Model License's $20 million API threshold carefully before building a hosted service on top of it.

Written byTheo Park

Theo Park runs the AI desk at Pandromeda. He follows model launches from the frontier labs and the open-weight community, tracks the assistants and developer tools built on them, and explains what each release changes on pricing, capability and safety. His reporting leans on primary sources: model cards, technical reports, API documentation and the companies' own announcements.

More from Theo ParkAll AI coverage