Playbooks · Coding tasks
Judge Muse Spark Open before you download it: 816B weights, a custom license, and a cluster-sized hardware bar
Written November 1, 2026 · Edited November 4, 2026
For a builder tempted by free 816B open weights: the license strings, the gated form, the hardware that can actually hold the model, and whether the hosted route is cheaper than the cluster, judged from the model card and the license text.
What changed
Meta Superintelligence Labs released Muse Spark Open on 15 October 2026. The model card describes an 816B total, 42B active parameter mixture-of-experts model, with 256 experts and 8 active per token, a 1M-token context at inference, and text, image, video, and audio input with text output. It is built for coding and agentic work, with extra safety post-training for use as a standalone model without system guardrails, and the card recommends deploying it as part of a broader system with guardrails suited to the use case. The weights are available in BF16 and NVFP4 versions. The download is gated: the access form asks for first and last name, date of birth, country, affiliation, and job title, and its heading asks for a full legal name, date of birth, and full organization name with all corporate identifiers, warning that not following the instructions may prevent access to this model and others on Hugging Face. Call it open-weight, not open source: the weights are downloadable, but they come under a custom license with conditions that shape what you can ship.
Strengths
On the model card's own benchmark table, run at xhigh reasoning effort with temperature 1.0, top_p 1.0, and top_k disabled, Muse Spark Open leads the compared open-weight models on several agentic measures: JobBench 61.8, against 61.4 for GLM 5.3 and 54.3 for Kimi K3; MCP-Atlas 89.2, against 84.5 for Qwen3.8-2.4T-A42B; HLE-Diamond without tools 26.1, against 22.2 for Kimi K3; and ZeroBench with tools 50.0, against 41.0 for Kimi K3. On safety measures in the same table, where lower is better, it records 0.5 on StrongREJECT v2, against 62.9 for GLM 5.3 and 66.6 for Kimi K3. The card also records a 98.1% honesty rate on the MASK benchmark and a 0% misbehavior rate on the Agentic Misalignment benchmark, and states that the model refused 99.2% of jailbreaks on an internal benchmark. These are Meta's own figures for the released checkpoint under the tested conditions. The card also gives the model a wide operating range: reasoning effort can be set to low, medium, high, or xhigh, with xhigh the default, and the supported frameworks named on the card are SGLang, vLLM, Hugging Face, TensorRT-LLM, Unsloth, and llama.cpp, with the model also optimized to run on two Mac Studios with M5 Ultra using MLX.
Weaknesses and limits
First, the license. The Muse Model License requires the license and usage policy to be included in copies and derivatives, and requires "Built with Muse Spark Open" to be prominently displayed in the user interface of each product or service that uses the model or any derivative, or in its documentation, website, or settings menu. "Derivatives" includes any AI model created, trained, modified, or otherwise improved using the model or any outputs or results of it, so a model trained on its outputs carries the notice too. If you or your affiliates provide access to AI models as a product or service through an API or similar interface, and aggregate gross revenue exceeds $20 million in any consecutive 12-month period, you must notify Meta and enter a separate written agreement with Meta before offering the model or derivatives as part of that model-as-a-service business. That clause does not apply to internal use, research, other commercial activity that is not model as a service, or use of the model as a component powering features of your own product or service that is not model as a service. Meta claims no ownership of outputs. The attached usage policy states the model is not intended for individuals under the age of 18, and lists prohibited uses. Read the license and the policy in full before building on the model: the summary here is not a substitute for them.
Second, the hardware bar. The card states that running the model on your own system requires a GPU with at least 700 GB of VRAM for NVFP4 and 2 TB for BF16. The 42B active figure describes the parameters used per token; all 816B weights still have to be available to the serving system, plus memory for context, activations, and runtime overhead. A community GGUF build from ggml-org lists smaller files, but it is not Meta's checkpoint: BF16 1.64 TB, Q8_0 872 GB, Q4_K_S 470 GB, and Q2_K 304 GB, described as fitting in 512 GB of unified memory, plus a separate vision and audio encoder file. Smaller files come from quantization, and the card warns that quantized inference may show minor quality differences in edge cases compared with full precision. Do not treat a GGUF file size as the card's VRAM requirement.
Third, the benchmarks cut both ways. On the card's table, Muse Spark Open trails the leaders on the overall Artificial Analysis Intelligence Index, 39 against 45 for GLM 5.3, 44 for Kimi K3, and 40 for Qwen3.8-2.4T-A42B, and trails on coding measures: DeepSWE 1.1, 60 against 68 for Kimi K3 and 67 for GLM 5.3; Terminal-Bench 2.1, 76.4 against 85.0 for Kimi K3 and 83.9 for GLM 5.3. The card's own limitations section states that the model may produce inaccurate, biased, or objectionable responses, may still make errors in multi-step reasoning in novel scenarios, has not been evaluated on all languages in its training data, and that results for the released checkpoint should not be assumed to apply to modified models.
What to do
- Read the license before the download. Read sections 2 and 3 and the usage policy in the license text linked below, and decide whether the attribution notice and the model-as-a-service clause fit how you plan to ship. If you plan to train another model on its outputs, the derivative definition applies to that model too.
- Decide whether to pass the gate. The access form asks for your legal name, date of birth, and organization details, and you cannot edit the form after submission. If you are not willing to share those details with Meta under its privacy policy, do not start the download: use a hosted provider instead, and read that provider's own terms.
- Price the hosted route on the provider's own listing before comparing it with hardware. The card states the model is available through third-party inference providers, and gives no prices. This page gives no hosted price, because no provider listing was read for it. Open the listing of the provider you would actually use, note its input, cached input, and output prices and its context limit, and price one month of your own token volumes from those figures. Compare that total with the hardware in step 4, not with the word "free".
- Match hardware to the card's figures, not to the active count. For self-hosting, budget at least 700 GB of VRAM for NVFP4 or 2 TB for BF16, plus the context and runtime overhead, on a framework the card names. If your hardware cannot meet the figure for the build you want, the hosted route or a quantized build is the honest alternative, and a quantized build needs its own quality check on your tasks.
- Benchmark it on your own tasks at xhigh before adopting it. Run three to five tasks that look like your real work, at xhigh reasoning effort, the setting the card's benchmarks used, and score them yourself. Meta's table is a starting point for that comparison, not a result for your workload. If you ship it in a product, add the attribution notice and the guardrails the card recommends, including human confirmation for irreversible actions in agentic use.
Limits of this evaluation
- No run of the model was made for this page. Every figure here comes from the model card, the license text, or the GGUF file listing named in the sources, with the conditions the card states.
- Benchmark figures are Meta's own, for the released checkpoint at xhigh reasoning effort. They do not cover every downstream configuration, and the card states they should not be assumed to apply to modified models.
- The GGUF figures come from a community build, not from Meta. They describe file sizes, not the card's VRAM requirements, and the two must not be mixed.
Related playbooks
- Run Muse Spark in another harness via the Model API
- Check whether your session is really on Muse Wonder or Wonder Full — and which route you are paying through
- MAI-Code-1.1-Flash: local coding on RTX Spark
- Mistral Large 4 Le Chonk: a trillion parameters, open
Sources
- https://huggingface.co/meta-models/Muse-Spark-Open-816B-A42B/blob/main/README.md
- https://huggingface.co/meta-models/Muse-Spark-Open-816B-A42B/blob/main/LICENSE.md
- https://huggingface.co/ggml-org/Muse-Spark-Open-816B-A42B-GGUF
Back to: Coding playbooks hub · Also: Run Liquid AI's large 24-billion-parameter agent model on your own computer — if your memory fits it · Run Liquid's LFM3-1.7B privately on your own machine: what runs, on what, and what it is not for · Pick the cheapest model for the job