Meta has released Muse Spark Open, an open-weight mixture-of-experts model with roughly 816 billion total parameters, about 42 billion active per token, and a one-million-token context window. The Hugging Face launch guide describes it as "the most capable US open-weights model to date," a superlative that Meta's own benchmark table supports only for some tasks (see below). It reads text, images, audio and video, writes text, and was trained for agentic coding: reading a repository, calling tools, and fixing its own mistakes over long sessions.
For anyone following Meta's Muse line, this is the full-size sibling of Muse Glimmer, the 30B open model from August, and the open counterpart to the hosted Muse Wonder and Muse Code products. Apple has already shown it running across four Mac Studios in its MLX and RDMA research.
This post sticks to what the Hugging Face guide, NVIDIA's technical blog and AWS's SageMaker walkthrough actually say, and flags what is still unverified.
Meta's own announcement, as posted on X:
Update (October 24, 2026): a "150M tokens per MW" claim, checked
A story circulating on October 23 says Muse Spark Open "reaches 150M tokens per MW" on Nvidia hardware. We could not trace that figure to a primary source. Nvidia's inference page makes general claims (up to 50x throughput per megawatt versus Hopper on agentic workloads, citing InferenceX) but, in our search, nothing specific to Muse Spark Open.
The closest primary data we found is SemiAnalysis's public InferenceX comparison of GB300 NVL72 and MI355X on Muse Spark Open. When we opened it, the page showed GB300 NVL72 at about 18.3M, 14.1M and 9.0M tokens per second per megawatt at 134, 195 and 256 tokens/s/user (FP4), against 7.9M, 4.4M and 2.6M for MI355X. A search snippet of the same page earlier showed different interactivity points, so the page is live and changes between runs. None of these values is near 150M.
Possible explanations are a different unit (tokens per megawatt-hour, or a different batch and interactivity target) or an error in the headline. Until Nvidia or Meta publishes the setup, treat 150M as unconfirmed, and check the InferenceX page's date selector before quoting any number.
TL;DR: the questions people will ask
| Question | Answer |
|---|---|
| Size? | About 816B total, about 42B active per token |
| Architecture? | Mixture of experts, 256 experts per layer, 8 chosen per token, 62 layers |
| Context? | Up to 1,048,576 tokens |
| Inputs and outputs? | Text, images, audio, video in; text out |
| Checkpoints? | BF16 (about 1.5 TB) and NVFP4 (4-bit, for Blackwell GPUs) |
| Licence? | Muse Model License: commercial use allowed, "Built with Muse Spark Open" attribution, separate deal for MaaS above $20M revenue |
| Can reasoning be disabled? | No; effort levels low, medium, high, xhigh (default xhigh) |
| Cheapest way to try it? | A 2-bit llama.cpp quant on a 512GB Mac, or SageMaker JumpStart |
| Benchmarks? | Meta table: leads MCP-Atlas (89.2), trails Kimi K3 on Terminal-Bench 2.1 (76.4 vs 85.0). No independent runs yet |
What is actually in the model?
Mixture of experts routing diagram showing a token sent to a few of many experts in Muse Spark Open
According to the Hugging Face launch post, the language model has 62 layers and a hidden size of 8192, with 128 query heads and 16 key-value heads (grouped-query attention). Every layer replaces the dense feed-forward block used in Muse Glimmer with an MoE block of 256 experts, and a router picks 8 per token. The experts hold about 97 percent of all parameters.
That has a practical consequence the post spells out: all 256 experts must sit in memory, so you need the memory of an 816B model, but each token costs the compute of roughly a 42B model. Fast and big, but not small.
Other details from the same post:
- A 1.8B-parameter perception encoder turns images and video frames into tokens, using 14 by 14 patches.
- Audio goes through a separate convolutional encoder that emits one token every 80 ms, or 12.5 tokens per second.
- Input and output embedding matrices are separate, about 1.66B parameters each.
- The chat template uses a channel-based format first seen in gpt-oss, where the model routes reasoning, commentary and the final answer to different recipients.
- Both Muse Glimmer and Muse Spark Open share a tokenizer and vocabulary.
How much memory does a 1M-token context cost?
Hugging Face does the arithmetic: grouped-query attention adds 126,976 values per token to the KV cache, about 250 KB per token in BF16, or roughly 266 GB for a full million-token window. So the 1M context is real, but filling it is a serious memory commitment on top of the weights. Plan capacity for the cache, not just the checkpoint.
What was it built to do?
Meta describes the model, in the words relayed by AWS and NVIDIA, as trained on real developer workflows: multi-file refactors, long debugging sessions, tool calling and first-attempt accuracy on large codebases. NVIDIA adds that it uses goal conditioning and context compaction to stay on track in long sessions, browses the web with images, spawns subagents for decomposable work, and does computer use across desktop, mobile and browser. It writes commentary before each action so a developer can see what the agent is doing.
NVIDIA's own demo, a video on its technical blog, shows the model diagnosing a bug from a screen recording with narrated audio. Treat that as a vendor demo, not an evaluation.
How do you run it?
Open-weights release illustration for running Muse Spark Open locally or on cloud GPUs
Transformers
The Hugging Face guide uses AutoModelForMultimodalLM with a DistributedConfig that sets tensor and expert parallelism. The BF16 checkpoint needs four nodes of eight H200s (tensor parallel 32); NVFP4 fits one B200 node. The processor handles images, video (sampled at 2 FPS, capped at 96 frames by default) and audio (mono, 24 kHz).
llama.cpp
The easiest local route. Install the prebuilt binaries and serve a quant straight from Hugging Face:
curl -LsSf https://llama.app/install.sh | sh
llama serve -hf ggml-org/Muse-Spark-Open-816B-A42B-GGUF:Q2_K
Hugging Face reports that on a Mac M3 Ultra with 512GB, an imatrix-calibrated 2-bit quant ran at 18 tokens per second and used about 304 GB, leaving room for the KV cache. The built-in web UI accepts audio and images and exposes the reasoning lever. For a lighter-weight llama.cpp front end, see our look at Llama Windows.
SGLang, vLLM and NVIDIA
The reference SGLang configuration serves the NVFP4 checkpoint across four GB300 GPUs with the modelopt_fp4 quantization and a FlashInfer MoE runner. NVIDIA says out of the box on a GB300 NVL72 it saw up to 50K tokens per second per GPU and up to 600 tokens per second per user on agentic workloads, with speculative decoding, and offers a downloadable NIM container plus NeMo AutoModel (SFT and LoRA) and Megatron Bridge for post-training. Those are NVIDIA's numbers on NVIDIA hardware.
Amazon SageMaker JumpStart
AWS lists both variants in US East (N. Virginia) and US East (Ohio). NVFP4 runs on ml.p5e.48xlarge, BF16 on ml.p6-b300.48xlarge. It is a gated model, so you accept an EULA (accept_eula=True in the SDK). AWS notes the endpoint stays in your account, can sit in your VPC with KMS encryption and network isolation, and that you pay for the multi-GPU instance for as long as it runs. Deployment takes 15 to 20 minutes, per the AWS walkthrough.
Local clusters of Macs
Apple's research post reports about 36.7 tokens per second on four M5 Ultra Mac Studios; our write-up of it is linked above. The two data points (18 tokens per second on a 2-bit quant on one M3 Ultra versus roughly 37 on a four-machine 4-bit cluster) are not directly comparable because quantization, hardware and settings differ.
Is it really open? What the Muse Model License says
The weights are public, but they ship under the Muse Model License, not Apache 2.0 or MIT. The Hugging Face model page is gated: you log in and fill in a form before you can see the files. These are the terms as shown on the model page. Read the full text before a commercial launch.
| Term | What it says |
|---|---|
| Allowed | Use, copy, modify (including fine-tuning), distill, merge, distribute, sublicense and sell, under the conditions below |
| Attribution | Products or documentation must prominently display "Built with Muse Spark Open" |
| Derivatives | Defined broadly: models trained, distilled or improved using the software or its outputs |
| Model-as-a-Service | If you and your affiliates offer model access to third parties and pass $20 million in gross revenue over any 12 months, you must notify Meta and sign a separate agreement first |
| Exempt | Internal use, research, and embedding the model inside your own product |
| Outputs | Meta claims no ownership of outputs |
| Warranty | "As is." You are responsible for your derivatives |
| Usage Policy | Bans illegal activity, malware, weapons, harassment, impersonation, disinformation, unlicensed professional practice and circumventing safety measures. Not intended for users under 18 |
For a startup building an app on top, the practical impact is small: add the attribution line and you are fine. The clause that matters is the Model-as-a-Service threshold. If your business is reselling inference of this model at scale, Meta wants a deal. That is why "open-weight" is the accurate label. Meta also has a history of mixing open and closed releases in this family, which we covered in our look at what Meta got right with Muse.
What are the architecture details and recommended settings?
From the Hugging Face model card:
- Layers and experts: 62 layers, hidden dimension 8,192, 256 experts with 8 active per token.
- Attention: Grouped-query attention with 128 query heads and 16 key-value heads, gated attention with QK-norm. Vocabulary of 202,048 tokens.
- Vision and audio: A 1.8B-parameter Perception Encoder handles images, video and audio. Output is text only.
- Languages: Trained on more than 100 languages, but not evaluated on all of them.
- Sampling: Temperature 1.0, top_p 1.0, top_k disabled.
- Reasoning effort:
low,medium,highorxhigh(default). Usehighorxhighfor coding and agent work. Other values return an error. - Video: Very short or low-resolution clips (around 100 px) may be refused. Use at least 448 px on the short side.
- Memory: The card says about 700 GB of VRAM for NVFP4 and about 2 TB for BF16, which includes working memory beyond the raw checkpoint size.
What do the benchmarks say?
The earlier version of this post could not open Meta's benchmark table. The Hugging Face card now gives numbers, all at xhigh reasoning and all published by Meta:
| Benchmark | Muse Spark Open | Best rival listed |
|---|---|---|
| JobBench | 61.8 (highest listed) | Kimi K3: 54.3 |
| MCP-Atlas | 89.2 (highest listed) | Not stated |
| HLE-Diamond (no tools) | 26.1 (highest listed) | Not stated |
| ZeroBench (with tools) | 50.0 (highest listed) | Not stated |
| OSWorld Verified | 80.4 | Kimi K3: 84.8 |
| Terminal-Bench 2.1 | 76.4 | Kimi K3: 85.0 |
| GPQA Diamond | 89.0 | Qwen3.8: 93.5 |
| Artificial Analysis Intelligence Index | 39 | GLM 5.3: 45, Kimi K3: 44 |
The pattern is clear. It leads on agentic tool use (MCP-Atlas, JobBench) and some reasoning sets. It trails on terminal coding, desktop control and the aggregate index. Alexandr Wang called it "the most capable US open-weight model to date," and Meta concedes it trails some Chinese models on coding. The "US" qualifier is doing real work in that sentence. For a head-to-head with hosted frontier models see our Muse Wonder comparison.
What did Meta say about safety?
Meta recommends deploying the model inside a larger system with use-case guardrails. Reported results on the released checkpoint, with no system-level safeguards:
- Refused 99.2% of jailbreaks on an internal cybersecurity benchmark and 94.8% across aggregated public benchmarks.
- Refused over 98% of harmful requests on SecureBio's BioTIER benchmark.
- 98.1% honesty rate on MASK and 0% misbehavior on Agentic Misalignment.
These apply only to the released weights. Fine-tuning can degrade them, and the license makes you responsible for your derivatives. Meta advises human confirmation for irreversible agent actions and not exposing raw chain of thought to users.
Where can you call it today?
Meta for Developers lists hosted APIs at Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI. Weights are on Amazon SageMaker AI, Google Cloud Model Garden, Microsoft Foundry Models and Hugging Face. Supported runtimes are SGLang, vLLM, TensorRT-LLM, Transformers, Apple MLX, llama.cpp and Unsloth, with optimizations for AMD, Apple, Intel and NVIDIA. Fine-tuning works with PyTorch, TorchTitan, TRL and NeMo AutoModel.
LightSeek's TokenSpeed claimed day-0 support at 190 output tokens per second per user on NVIDIA B300 without speculative decoding, a vendor figure. One early reply asked Meta for an end-to-end fine-tuning example, which shows the gap: the weights landed before the recipes.
What is claimed and what is verified?
| Claim | Source | Status |
|---|---|---|
| 816B total, 42B active, 256 experts, 1M context | Hugging Face card, NVIDIA, AWS, Meta | Consistent across sources |
| "Most capable US open-weights model" | Meta, Alexandr Wang | Meta's own table supports it only for some tasks; it trails Chinese models on coding |
| Benchmark scores above | Hugging Face card | Meta-published, not independently reproduced |
| 50K tokens/s per GPU, 600 tokens/s per user | NVIDIA | Vendor numbers on GB300 NVL72 with speculative decoding |
| 18 tokens/s on M3 Ultra 512GB at 2-bit | Hugging Face | Reported by the launch partner, not independently reproduced |
| Helion kernels 1.03x to 3.95x faster in SGLang | Meta | Meta claim at kernel level |
When independent evaluations appear we will add an update section.
Which model for which job?
| Job | Reasonable pick | Why |
|---|---|---|
| Experiment on a laptop or single GPU | Muse Glimmer 30B | Small dense model from the same family |
| Private coding agent in your cloud account | Muse Spark Open NVFP4 on SageMaker | Weights in your account, VPC and KMS options |
| Maximum local capacity on Apple hardware | Muse Spark Open on 512GB Mac or a Mac cluster | See Apple's numbers |
| Hosted, no ops | Meta's Muse Code | No hardware to manage |
What this means for what you build or pay
If you run coding agents, an 816B open model changes the economics only if you can feed it memory. Per-token compute is modest, but the weights are large, and a full 1M context adds hundreds of gigabytes. For most teams the realistic paths are a managed endpoint in your own cloud account or a small GPU node using NVFP4. The privacy upside is real: prompts and source code stay on infrastructure you control. The risk is that agentic models can act on your machine, so sandbox them and limit tool permissions.
Related reading
-
Nvidia to invest $3B in Meta as the two expand their Muse collaboration
-
IQuest-Q1: a 320B open-weight MoE built for agentic CLI work
-
NVIDIA's DFlash 2 draft model for Muse Spark Open (speculative decoding, vLLM)
-
Muse Spark Open vs Muse Glimmer vs Nemotron 3.5 Bolt vs Super VL: hardware comparison
-
Ripple-65B Preview: a 65B dense open-weight model compared with Muse Spark Open
Specs and availability are accurate as of October 15, 2026 and may change as Meta, Hugging Face, NVIDIA and AWS update their pages.