Meta has released Muse Spark Open, making the weights of its flagship model for programming and AI agents publicly available. The model employs a mixture-of-experts architecture with a total of 816 billion parameters, but only approximately 42 billion parameters are activated per token. It supports a context length of up to 1 million tokens and accepts text, images, audio, and video inputs. Meta and Apple jointly optimized the MLX deployment: the official 4-bit quantized version can run on two M5 Ultra Mac Studio units, each equipped with 512GB of unified memory, achieving measured generation speeds of approximately 15.7 to 25 tokens per second across two machines, and up to 36.7 tokens per second with four machines. However, the claim of being the “most powerful open-weight model in the U.S.” is geographically specific: in Meta’s own comparison table, it still lags behind certain Chinese open models in overall benchmarks, code terminal tasks, and professional workload evaluations. Its license permits most research and commercial applications but includes usage restrictions, attribution requirements, and revenue thresholds for model services; therefore, it does not qualify as a fully open-source model in the conventional sense.Author and source: Meta
816 billion parameters does not mean that all 816 billion parameters are computed during each inference.
Muse Spark Open consists of 62 layers of language models and a perception encoder. Each mixture-of-experts layer contains 256 experts, with the router selecting 8 experts per token to participate in computation, resulting in a total of approximately 816 billion parameters and approximately 42 billion parameters activated per inference. About 97% of the parameters are stored within these expert networks.
This architecture reduces computational load without correspondingly alleviating memory pressure: during inference, although only a few experts are activated, all expert weights must remain resident in memory. BF16 checkpoints are approximately 1.5 TB; Meta’s released NVFP4 version compresses weights to about 493 GB, but when accounting for caches and intermediate tensors, a single 512 GB device still cannot accommodate the standard configuration. Meta’s model card specifies requirements for traditional GPU deployment: at least ~700 GB VRAM for NVFP4 and ~2 TB for BF16.
The model has a 1-million-token context window and was trained on over 100 languages. It can process text, images, videos, and audio, but its output remains text-only—it does not directly generate speech, images, or videos. Its 1.8B visual encoder converts images and video frames into tokens, while the audio encoder generates an audio token approximately every 80 milliseconds.
Meta is focusing on long-duration programming agents, including cross-file refactoring, continuous debugging, tool invocation, retrying after failures, and interacting with software based on screen visuals. The reasoning mode cannot be fully disabled; it can only be set to one of four intensity levels: low, medium, high, or xhigh. The default setting is xhigh, which consumes the most tokens.
"Desktop Run" is actually a computing cluster connected via Thunderbolt.
Meta emphasizes that this is the first time a model of its kind has been able to run in a desktop environment, but "desktop" here does not refer to an ordinary Mac.
The official setup uses two M5 Ultra Mac Studios, each with 512 GB of unified memory, connected via two Thunderbolt 5 cables and RDMA drivers. In pipeline parallelism, the first half of the model is placed on one machine and the second half on the other; in tensor parallelism, the computations for each layer are simultaneously distributed across multiple machines.
In tests with an 8192-token context followed by generation of 128 tokens, two Macs achieved a generation speed of 15.74 tokens per second with pipeline parallelism and 24.99 tokens per second with tensor parallelism; increasing to four machines raised the speed to 36.73 tokens per second. The corresponding first-token latency was approximately 8.4 seconds, 8.2 seconds, and 4.8 seconds, respectively.
(Apple technical article includes two sets of charts comparing first-token latency and generation speed when using two and four Mac Studios with pipeline and tensor parallelism; also includes diagrams illustrating the differences between model partitioning by depth and width.)
This configuration requires at least 1 TB of unified memory. According to The Wall Street Journal, the hardware cost exceeds $40,000. Therefore, it is more akin to a private inference workstation that can be placed in an office than a local model running on a typical personal computer.
However, the community’s aggressive quantization schemes have further lowered the barrier. The Hugging Face team tested three 1-bit quantized versions, each occupying approximately 344 GB and achieving around 21 tokens per second on a 512 GB M3 Ultra. The 2-bit version from llama.cpp occupies about 304 GB and was tested at approximately 18 tokens per second. The trade-off is that stronger compression may lead to accuracy loss in edge cases, and the context lengths used in these tests are significantly smaller than the advertised 1 million tokens.
The qualifier "strongest in the United States" is important.
Meta calls Muse Spark Open the "most powerful open-weight model to date in the U.S.," but does not claim it is the world's strongest open model.
On Meta’s Artificial Analysis Composite Index, Muse Spark Open scores 39, higher than Inkling and Nemotron 3 Ultra from the United States, but lower than Qwen3.8’s 40, Kimi K3’s 44, and GLM 5.3’s 45.
Its strengths are primarily concentrated in certain Agent tasks: JobBench scores 61.8, outperforming other open models in the table; MCP-Atlas scores 89.2, also leading. It also performs well on tool-free evaluations such as HLE-Diamond, SimpleVQA, and physical understanding.
However, its software engineering capabilities are not superior across the board. Muse Spark Open scores 60 on DeepSWE 1.1, below GLM 5.3’s 67 and Kimi K3’s 68; on Terminal‑Bench 2.1, it scores 76.4, while Qwen, GLM, and Kimi achieve 82.0, 83.9, and 85.0 respectively. In the professional knowledge task GDPVal-AA v2, its Elo rating of 1401 is also lower than that of several competitors.
(The model card here includes a comprehensive benchmark table covering general capabilities, agent performance, scientific reasoning, coding, multimodal, and security evaluations.)
Therefore, a more accurate conclusion is that Muse Spark Open significantly narrows the gap between U.S. open-weight models and Chinese open models, establishing an advantage in tool calling and certain visual agent tasks, but has not yet achieved global leadership in comprehensive reasoning and programming benchmarks.
These figures are not entirely consistent or independently benchmarked. Meta notes that some results come from official leaderboards, some were run by Meta itself, and scores for models like Beam and Mistral were self-reported by developers; the agent tools and system prompts used for third-party models may not have been individually optimized for each model.
The weights are open, not the full training process.
Muse Spark Open provides both BF16 and NVFP4 checkpoints and is supported by Transformers, vLLM, SGLang, TensorRT-LLM, llama.cpp, Unsloth, and MLX. Developers can download, modify, fine-tune, distill, and redistribute the model, as well as integrate it into commercial products.
However, the Muse Model License includes several important conditions: the product or documentation must prominently display “Built with Muse Spark Open”; if a business operates a “model-as-a-service” offering via API or other means, and the business and its affiliates generate more than $20 million in total revenue over a consecutive 12-month period, a separate agreement with Meta is required before providing Muse Spark Open or its derivative models. The license also prohibits use for military, weapons, deception, hazardous infrastructure, and similar purposes.
Therefore, "open weights" is more accurate than "open source." Meta has not released the full pre-training code, data catalog, or reproducible training recipe. The model card provides only a general overview, stating that the training materials include publicly available sources, web-scraped data, third-party data, data derived from Meta products and services, and synthetic data, curated by external vendors and Meta employees.
The collaboration between Meta and Apple is more noteworthy than individual benchmark scores.
Meta and Apple have long clashed over advertising tracking, privacy policies, and software and hardware platforms, yet they have now jointly completed adaptations for the model with MLX, unified memory, and Thunderbolt RDMA. Apple needs to demonstrate that its high-end Macs are not only suitable for developing small models but can also handle private AI models approaching trillion-parameter scales. Meta, on the other hand, requires a more intuitive local deployment reference than traditional multi-GPU servers.
The two companies demonstrated not that "large models can now run on everyone's laptop," but rather an alternative infrastructure option: running state-of-the-art models on two to four low-power desktop workstations with high unified memory, without sending code, videos, or enterprise data to the cloud. For teams requiring sustained model usage and prioritizing data residency, this one-time hardware investment may compete with ongoing cloud GPU rentals.