Meta Releases Muse Spark Open: 816B Open-Weight MoE With 1M Context

Meta says Muse Spark Open is the most capable US open-weight model to date. Weights ship in BF16 and NVFP4 on Hugging Face. The open version received extra safety post-training, which Meta says makes it more robust without system-level guardrails.

The model card lists 62 layers and a hidden dimension of 8192. Each MoE layer holds 256 experts, and top-k routing activates 8 experts per token. Attention is grouped-query, with 128 query heads and 16 KV heads. A separate 1.8-billion-parameter Perception Encoder handles visual input. The vocabulary has 202,048 tokens, and training data covers more than 100 languages. Reasoning depth is set with reasoning_effort: low, medium, high or xhigh, with xhigh as the default.

Meta compares Muse Spark Open with Kimi K3, GLM 5.3, Qwen3.8-2.4T-A95B and others at maximum reasoning effort. Its strongest results are on tool use and knowledge reliability. MCP-Atlas scores 89.2, ahead of Qwen3.8 at 84.5 and GLM 5.3 at 84.2. JobBench scores 61.8, narrowly ahead of GLM 5.3 at 61.4. HLE-Diamond without tools scores 26.1, versus 22.2 for Kimi K3. AA Omniscience scores 25, the top score in Meta's table. ZeroBench with tools scores 50.0, versus 41.0 for Kimi K3.

Coding is the weak spot. On Terminal-Bench 2.1 with the Terminus 2 harness, Muse Spark Open scores 76.4. Kimi K3 reaches 85.0, GLM 5.3 scores 83.9 and Qwen3.8 scores 82.0. Its Artificial Analysis Intelligence Index is 39, behind GLM 5.3 at 45, Kimi K3 at 44 and Qwen3.8 at 40. Meta's claim of the most capable US open-weight model holds in its comparison, but Chinese open models still lead in the AA Intelligence Index. Some rival scores are self-reported, and the evaluation methodology lists each source.

Per Meta's developer guide, BF16 weights are about 1.6 TB. NVFP4 quantizes only the experts and lands near 530 GB. The model card recommends at least 700 GB of VRAM for NVFP4 and 2 TB for BF16. In data centers, NVFP4 runs on 4 Blackwell Ultra GPUs or 8 H200s. On desktop, it runs on 2 M5 Ultra Mac Studios with 512 GB each, linked over Thunderbolt 5 with RDMA; MLX tensor parallelism splits every layer across both machines. For workstations, Unsloth Dynamic GGUFs down to 2-bit fit in about 300 GB of combined memory.

Day-one engines include SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, MLX, llama.cpp and Unsloth. Hosted APIs are live on Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI. OpenRouter lists $1.25 per 1 million input tokens and $4.25 per 1 million output tokens. Two draft models speed up decoding: DFlash V2 for SGLang and DSpark for vLLM. Meta reports 3.7x single-request throughput with DFlash V2 on AMD MI355X. Fine-tuning works with TorchTitan, Hugging Face TRL and NVIDIA NeMo AutoModel.

The Muse Model License permits use, fine-tuning, distillation and redistribution. Two conditions matter most: products must display “Built with Muse Spark Open,” and model-as-a-service providers above $20 million in 12-month revenue need a separate agreement with Meta. Internal use, research and powering a company's own product features are exempt from the revenue clause.

In the comparison table, Muse Spark Open has 816B total parameters and 42B active parameters. Kimi K3 has 2.8T total and 104B active, GLM 5.3 has 753B total and about 40B active, and Qwen3.8-2.4T-A95B has 2.4T total and 95B active. Muse Spark Open supports a 1M-token context, while Kimi K3 and GLM 5.3 list 1,048,576 tokens and Qwen3.8 lists 262,144 native, up to 1,010,000. Muse Spark Open takes text, image, video and audio; Kimi K3 takes text, image and video; GLM 5.3 and Qwen3.8 take text. Released weights are BF16 and NVFP4 for Muse Spark Open, MXFP4 with QAT for Kimi K3, BF16 and FP8 for GLM 5.3, and BF16 for Qwen3.8. Vendor-reported Toolathlon Verified scores are 72.8 for Muse Spark Open, 76.5 for Kimi K3, 73.0 for GLM 5.3 and 72.5 for Qwen3.8. Vendor-reported HLE with tools scores are 61.0, 56.0, 62.5 and 56.2, respectively. Vendor-reported DeepSWE 1.1 scores on their own harness are 60, 67.5, 66.9 and 56.6. Only Muse Spark Open's desktop setup is disclosed: 2 M5 Ultra Mac Studios.