← Back to feed

Meta released the weights for Muse Spark Open on October 15, six weeks after launching Muse Spark 1.3 as API-only with open weights described as uncertain. The release makes Muse Spark Open the largest open-weight model from a US lab and the first in its parameter class to run entirely on consumer-grade Apple silicon.

What Shipped

Muse Spark Open is a sparse Mixture-of-Experts model at 816 billion total parameters and approximately 42 billion active per token. The architecture uses 256 experts per MoE layer with 8 selected for each token, 62 transformer layers, and 8,192 hidden dimensions with grouped-query attention (128 query heads, 16 KV heads). The context window is 1 million tokens.

The model accepts text, images, audio, and video as input. The BF16 checkpoint is roughly 1.5 TB of safetensors. An NVFP4 build is available for NVIDIA Blackwell hardware, where it fits on a single B200 node. BF16 requires tensor parallelism across 32 GPUs — four nodes of eight H200s each.

The Apple Silicon Angle

Meta specified that Muse Spark Open was optimized to run on two Apple M5 Ultra Mac Studios. M5 Ultra’s unified memory architecture provides 512 GB per machine, giving two units a combined terabyte of addressable memory. The 1.2 TB/s memory bandwidth is sufficient to read the active parameter slice per token.

This makes Muse Spark Open the first model at this scale explicitly validated for desktop-class hardware, though the two-Mac configuration is not a consumer setup. At roughly $10,000 for two M5 Ultra Mac Studios, it is accessible for labs and well-funded developers without H100 clusters.

Helion Kernels and Performance

The release ships with Helion GPU kernels — Meta’s PyTorch-native kernel authoring language — replacing custom CUDA implementations for operations including RoPE, RMSNorm, and quantization. The published speedups are 1.03–3.95x over CUDA-specific kernels on NVIDIA GB300 hardware, with equivalent gains on AMD MI350 (3% end-to-end throughput improvement). The kernels are auto-tunable across hardware targets, so the same code path runs on NVIDIA, AMD, and Intel accelerators.

The License

Meta released Muse Spark Open under the Muse Model License with limited commercial terms. The license allows many developers to use the model for free; specific thresholds for commercial restrictions are in the license text. This is tighter than the Llama license terms Meta used historically, but consistent with how the company has handled the Muse model series.

Inference Partners

The weights are available on Hugging Face. Hosted inference is live on Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, and Together AI. Enterprise integrations include Amazon SageMaker AI, Google Cloud Model Garden, and Microsoft Foundry Models. Day-zero support in SGLang and vLLM is available, and the model is listed on Hugging Face’s Transformers library.

Context

Muse Spark 1.3 launched September 3 with open weights described as uncertain. Stack Futures noted at the time that the efficiency gains — 20% fewer tool calls, 25% fewer tokens per task — were measurable, but the open-source commitment that built the Llama brand was in question. The October 15 delivery answers that question.

Muse Spark Open shares the same base model as Muse Spark 1.3. The open-weight version received additional safety post-training, per Meta, to make it more robust as a standalone model operating without system guardrails.

The model has additional safety post-training beyond Muse Spark 1.3, evaluated under Meta’s Superintelligence Scaling Framework. The full evaluation report is forthcoming; results passed the thresholds Meta’s framework requires for release.

Benchmarks

Meta has not published a comprehensive SWE-bench Verified or tau-bench figure for Muse Spark Open as of release. The model shares the same base weights and pretraining as Muse Spark 1.3, which scored 64.1 on LiveBench Agentic Coding in the September 2026 run under xHigh effort. Independent benchmark results are expected once the model is available across inference providers.