Meta has released Muse Spark Open, a large open-weight AI model built for agentic coding that also accepts images, audio and video. The New Stack frames it as Meta's answer to the best open-weight models coming out of China. One detail decides what this means for you, and it comes from Hugging Face's launch post. The model holds about 816 billion parameters but uses only about 42 billion for each token, and all 816 billion have to stay in memory. It runs at the speed of a much smaller model but needs the memory of a giant one.
The background you need
Open-weight. A model's weights are the learned numbers that make it work. An open-weight model lets you download those numbers and run the model on your own hardware. You are not limited to calling it through a company's API. Open-weight does not automatically mean open source. The training data and code may stay private, and the license can still restrict how you use the model.
Parameters. Parameters are those learned numbers, counted. More parameters usually means more capability. It also means more memory to load the model and more computation to run it.
Dense versus mixture-of-experts (MoE). In a dense model, every parameter takes part in producing every token. A mixture-of-experts model splits part of each layer into many smaller sub-networks called experts. A small component called a router picks a few experts for each token, and only those experts run.
An analogy: picture a hospital with 256 specialists, where each patient sees 8 of them. Each visit is quick because only 8 doctors do any work. The hospital still has to employ, house and pay all 256. That is the trade-off here. Compute per token is small, but memory covers the whole staff. The analogy breaks in two places. First, the router picks experts separately for every token in every layer, not once per question. Second, the experts are not specialists in human topics like Python or medicine. What each one learns is whatever pattern training happened to push toward it.
Precision and quantization. Each parameter is stored as a number with some precision. BF16 uses 16 bits per number and is a common full-quality format. Quantization stores the numbers in fewer bits, such as 4. The file gets much smaller, at some risk to quality. NVFP4 is a 4-bit format designed for NVIDIA's Blackwell generation of GPUs.
Context window and KV cache. The context window is how much input the model can consider at once, measured in tokens (chunks of text, or pieces of images and audio). While it works, the model keeps a running memory of earlier tokens called the KV cache. That cache grows with every token. So a long context costs memory on top of the weights themselves.
Agentic coding. This means the model does more than suggest code. It reads a codebase, makes changes, runs tools, looks at the results and corrects itself over many steps.
What actually changed
Some of this was already true before the launch. According to Hugging Face, Meta released Muse Glimmer in August. Glimmer is a small dense model in the same Muse series, which Hugging Face describes as ideal for running locally. Muse Spark Open uses the same tokenizer and vocabulary as Glimmer, and the same image encoder design.
Here is what is new, according to Hugging Face:
- MoE at large scale. Each layer has 256 experts, and the router selects 8 per token. The experts hold about 97% of the model's parameters. The total is about 816B, with about 42B active per token.
- Audio input. A new audio encoder turns sound into 12.5 tokens per second. This lets the model handle video together with its soundtrack, which Hugging Face describes as an upgrade from Glimmer.
- A one-million-token context window. Text, image, video and audio tokens all share it.
- Two checkpoints. One is in BF16, at about 1.5 TB. The other is NVFP4, which Hugging Face says has negligible loss in accuracy.
- Reasoning that cannot be switched off. You choose an effort level of low, medium, high or xhigh. The code example in the post shows xhigh as the default.
- A channel-based chat template. This format is called Harmony, which Hugging Face says was first introduced by gpt-oss.
Hugging Face calls Muse Spark Open "the most capable US open-weights model to date." The New Stack's headline positions it against China's strongest open-weight models. Neither source, in the material available here, gives benchmark numbers that back up the comparison.
What it means for developers who don't track every launch
The MoE trade-off is the detail that matters most, because it decides where you can actually use this model.
It is not a laptop model. Hugging Face states that deployment needs the memory of an 816B-parameter model. The BF16 files alone are about 1.5 TB. The setups the post demonstrates in the transformers library are datacenter hardware. The BF16 checkpoint is shown across 4 nodes of 8 H200 GPUs each. The NVFP4 checkpoint is shown on a single B200 node. These are configurations Hugging Face describes, not official minimum requirements. Still, they tell you the scale involved. For local work, Glimmer remains the model Hugging Face points to.
Speed could make it practical to host. Each token only touches about 42B parameters, so Hugging Face says it runs at the much faster speed of a 42B model. That could make it attractive for inference providers. For most readers, the realistic way to try it is a hosted endpoint or a team cluster, not your own machine.
Long context costs extra memory. Hugging Face calculates the KV cache at about 250 KB per token in BF16. That comes to about 266 GB for a full one-million-token context, on top of the weights. If you plan to feed it a whole repository, budget for that.
Reasoning tokens add up. You cannot turn reasoning off. If you pay per token, set effort to low for simple tasks.
Odd output may be a template problem. If tool calls come back garbled, Hugging Face suggests checking which Harmony channel the model routed its answer to.
The 42B figure describes compute per token, not memory. Hugging Face says all 256 experts stay loaded, so plan your memory for the full 816B parameters, plus the KV cache if you use long contexts. Also note that reasoning cannot be disabled, so use the low effort setting when you want fewer tokens on easy tasks.
What to watch next: which hosted providers offer the model and at what price. Also watch whether community quantizations through llama.cpp, which Hugging Face says has deployment recipes alongside vLLM and SGLang, bring it onto smaller setups.
What we don't know yet
- License terms. Neither source, as available here, spells out what the license allows for commercial use.
- Benchmarks. There are no head-to-head numbers against the Chinese open-weight models The New Stack's headline refers to, or against closed models.
- Minimum hardware. Hugging Face describes tested configurations. It does not give the smallest setup that works, or how well heavily quantized versions perform.
- NVFP4 quality in practice. Hugging Face describes the accuracy loss as negligible. Independent testing on coding tasks has not been shown yet.
- Hosted availability and pricing. Neither source says which providers will serve it or what it will cost.
Originally published at pickuma.com. Subscribe to the RSS or follow @pickuma.bsky.social for new reviews.
Top comments (0)