Welcome Muse Spark Open: open-weight frontier multimodal mixture-of-experts model from Meta
Meta has published Muse Spark Open, an open-weight multimodal mixture-of-experts model, alongside a Hugging Face post covering how to run it, fine-tune it, and what was built with it. The full-size member of the Muse series follows Muse Glimmer, which Meta released in August and which is a small dense model; both share a tokenizer and vocabulary, and the chat template follows Muse Glimmer's with additions for the new modalities.
The headline numbers: approximately 816 billion total parameters with around 42 billion active per token, and a context window of up to one million tokens. The configuration sets max_position_embeddings to 1,048,576, and that single window holds image, video and audio tokens alongside text. Meta describes the model as agentic and coding-optimized, trained on real developer workflows such as multi-file refactors and long debugging sessions, able to plan, act, observe and self-correct over long tasks, including browsing the web with images. Because it also accepts images, audio and video, the same model can read a screenshot, listen to a voice note, or watch a screen recording while working on code. Meta calls it the most capable US open-weights model to date. Checkpoints are published as Muse-Spark-Open-816B-A42B and Muse-Spark-Open-816B-A42B-NVFP4.
Architecturally, the language model has 62 layers and a hidden dimension of 8192. Each layer uses 128 query heads and 16 key-value heads with grouped-query attention, each head 64 dimensions, plus an output gate and output normalization in the attention block. Normalization is applied to queries and keys (QK-norm), and each block combines input and output through a learned residual gate rather than a plain sum. Where Muse Glimmer has one dense feed-forward network per layer, Muse Spark Open substitutes a mixture-of-experts block with 256 experts, of which a router selects 8 per token. Each expert projects from 8192 to 4096 dimensions and back, then applies its own normalization. The experts hold roughly 97% of total parameters. The input embedding and output layer are separate matrices of about 1.66B parameters each.
The practical consequence of that routing is a split between memory and compute. All 256 experts must stay resident, so deployment needs the memory footprint of an 816B-parameter model, but each token only touches about 42B parameters, so it runs at the speed of a much smaller model. In practice this is the central operational constraint: the 816B figure describes what you must fit, while the 42B figure describes what you pay per token at runtime. The BF16 checkpoint is about 1.5 TB in safetensors format; in transformers it runs across 4 nodes of 8 H200 GPUs with tensor parallelism (TP=32). The NVFP4 checkpoint stores parameters in 4-bit precision, which Meta says has negligible accuracy loss, and runs on a single B200 node on NVIDIA Blackwell GPUs. Deployment recipes are available for llama.cpp, vLLM and SGLang.
Perception runs through a 1.8B-parameter encoder with 50 layers, hidden dimension 1536 and 16 attention heads. It splits images into 14x14 patches and produces 4096-dimensional features, using learned absolute embeddings and 2D RoPE for position; each layer's feed-forward network uses LayerNorm and GELU. Every fourth layer and the last layer use full attention, with sparse attention elsewhere. Muse Glimmer 30B uses an encoder of the same architecture, and the design follows Meta's Perception Encoder work. Video goes through the same encoder as images.
Audio enters through a dedicated encoder: an audio tokenizer converts audio to tokens, the chat template marks positions with audio tokens, and the encoder emits one audio token every 80 ms, or 12.5 tokens per second, so ten seconds of audio becomes 125 input tokens. The encoder is a convolutional pyramid with stages of varying patch sizes and convolutional blocks of 1d kernels of size 7 and stride 1, repeated for patch sizes of 16, 6, 5 and 4, ending in RMSNorm, linear projection and RMSNorm. The processor expects mono audio at 24 kHz and pads waveforms to a multiple of 1,920 samples, with resampling handled by the load_audio utility (torchcodec, librosa or torchaudio backends).
The long context is enabled by grouped-query attention keeping the KV cache small relative to model size. Each token adds 2 x 16 heads x 64 dimensions x 62 layers, or 126,976 values, to the cache: about 250 KB per token in BF16, and roughly 266 GB for a full one-million-token context. That is a useful number to budget for even before expert weights enter the picture, and it suggests long-context serving will be dominated by cache memory as much as by parameters.
In transformers, the implementation exposes MuseSparkForConditionalGeneration for generation from multimodal inputs and MuseSparkProcessor for preparing conversations; the processor bundles the tokenizer, image and video processors, audio feature extractor and chat template. Parallelism is configured with DistributedConfig, where tp_size shards dense layers across devices and ep_size distributes experts. Reasoning effort is a lever mapping symbolic levels low, medium, high and xhigh to integers 16, 64, 256 and 512, defaulting to xhigh; a session cannot turn reasoning off, so Meta suggests keeping it at low for easier tasks to save tokens, and a numeric effort can be set through the system prompt. Images are expanded by the processor into the patch tokens the processed image requires, surrounded by image boundary tokens so text positions stay aligned with visual features, and multiple images per prompt are handled. Video prompts include frame timing and are sampled at 2 FPS by default, capped at 96 frames; the video processor's fps attribute can be raised, and passing load_audio_from_video=True enables audio in video. Higher frame rates improve fidelity but cost memory and prefill latency.
Tool calling works for text-only and multimodal inputs, with a channel-based Harmony chat template, the format first introduced by gpt-oss. The assistant chooses a recipient, with self for reasoning, commentary for user-visible remarks before and between tool calls, and user for the final answer; tool invocations are parsed from structured function-call blocks. The signal for builders is that inspecting routing channels is worthwhile when output quality drops.
Quantized local use is possible: Muse Spark Open ships in llama.cpp prebuilt binaries, installed via a shell script, and a GGUF repository offers quants. Serving a Q2_K quant is a one-liner with a quant suffix on the model ID, and the built-in WebUI at localhost:9931 accepts audio and images and exposes the reasoning lever. On a Mac M3 Ultra with 512 GB, Meta reports an imatrix-calibrated 2-bit quant running at 18 tokens/s with 304 GB used by the model, leaving room for KV cache. Meta provides no separate pricing because it is open weights.
Why it matters: developers and product teams get a frontier-scale open checkpoint that handles code plus image, audio and video through one interface, which lowers the barrier to building agents that watch screens, listen to recordings and edit repositories. The catch is deployment cost: the 816B memory footprint and the roughly 266 GB KV cache at full context mean self-hosting requires serious multi-GPU hardware, and 2-bit local quants trade accuracy for accessibility. Teams weighing open versus hosted models now have a much stronger open option, though the memory profile suggests most will consume it through quantized or hosted routes rather than full precision.
Based on reporting from the original publisher. Visit the source for full context and later updates.