On October 15, 2026, Meta released Muse Spark Open, an open-weights model for coding and agentic tasks: 816 billion parameters, of which about 42 billion are active per token, with a context window of up to 1 million tokens. You can download the weights, but “running it yourself” in practice means an eight-GPU H200 cluster or a Blackwell node, and for a Russian company it also means checking the license, access to the weights, and Meta’s status in Russia. For most companies, this is a reason to refresh their test suite and their own LLM roadmap, not to rush out and buy hardware.
Prepared by AIrassvet October 20, 2026 for IT directors, engineering leaders, and business owners evaluating LLMs on their own servers. We relied on the model card and technical breakdown on Hugging Face, a TechCrunch report, and a Pravo.ru report on the court ruling. We explain what the release means for a company considering an LLM on its own servers; a full review of all open models and step-by-step deployment instructions are not included in this article. We did not run the model ourselves: all quality and performance figures below come from Meta and Hugging Face, not our own measurements. Meta Platforms has been designated an extremist organization in Russia, and its activities are prohibited. This material is not legal advice.
Contents
- At a glance
- What Meta released on October 15
- What “open weights” means in the license
- What hardware you need to run it
- How strong the model is, according to Meta
- Restrictions for Russian companies
- Who should consider the model, and who should not
- What the company should do now
- How we checked the information
- Frequently asked questions
- Source documents
At a glance
- What was released. Muse Spark Open is Meta’s open-weights model: 816 billion parameters, about 42 billion active per token, a context window of up to 1,048,576 tokens, with text, images, audio, and video as inputs and text as output.
- Hardware. According to Hugging Face, the compressed NVFP4 version runs on eight H200s with 141 GB each at full context, or on eight RTX Pro 6000s with 98 GB each at a context length of 131,072 tokens. The full BF16 weights take about 1.5 TB.
- License. The Muse Model License allows commercial use and fine-tuning, but requires the line “Built with Muse Spark Open.” A separate agreement with Meta is required if you sell access to the model to third parties as a service and your company’s combined revenue with affiliates exceeds $20 million over 12 months.
- Quality. In Meta’s table, the model scores 39 on the Artificial Analysis Intelligence Index. In the same table, GLM 5.3 scores 45 and Kimi K3 scores 44. Meta calls it the strongest American open model, but it is not the top open model overall.
- Russian language. Meta says the model was trained on data in more than 100 languages, but not all were evaluated. Performance in Russian has not been published and must be tested on your own use cases.
What Meta released on October 15
Meta published Muse Spark Open on October 15, 2026, and the model card and Hugging Face breakdown appeared that same day. TechCrunch says Meta calls it the strongest American open-weights model. The model is designed for advanced coding and agentic work. Key details from the Hugging Face technical breakdown:
| Parameter | Value |
|---|---|
| Architecture | Mixture-of-Experts: 256 experts per layer, 8 selected per token |
| Parameters | about 816 billion total, about 42 billion active per token |
| Context | up to 1,048,576 tokens |
| Input | text, images, audio, video |
| Output | text only |
| Reasoning modes | low, medium, high, xhigh; xhigh is the default and cannot be turned off |
| Weight formats | BF16, NVFP4, 3- and 4-bit MLX, GGUF for llama.cpp (Q2_K, Q4_K_S, Q8) |
| Support | Transformers, vLLM, SGLang, llama.cpp, MLX |
Mixture-of-Experts (MoE) is an architecture in which the model is made up of many “experts,” and only a subset of them is used for each token. That means roughly 42 billion parameters are computed at each step, but all 816 billion must be kept in memory. This is the model’s main business implication: compute requirements are moderate, but memory requirements are very high.
What “open weights” means in the license
Open weights means the model files can be downloaded and run on your own hardware. That is not the same as an open license: usage terms are set by the Muse Model License. According to the model card as of October 20, 2026, the main terms are:
- Allowed: use, copy, modify, and fine-tune the model, create derivative models, and distribute and sell them subject to the terms.
- Required attribution. The product or documentation must clearly include the phrase “Built with Muse Spark Open.”
- Threshold for using the model as a service. If you provide third parties with access to the model through an API or similar interface, and your company’s combined revenue with affiliates exceeds $20 million in any rolling 12-month period, you must notify Meta and sign a separate agreement before launching such a service. Internal use, research, and features inside your own product that does not sell access to the model are not covered by this condition.
- Outputs. Meta does not claim rights to the model’s answers.
- Usage policy. Appendix A prohibits, among other things, illegal activity, malicious code, privacy violations, and disinformation. The model is not intended for people under 18. In addition, you may not violate applicable laws, including trade restriction laws.
- Access to the weights. The repository is subject to manual moderation: the form requires your full legal name, date of birth, and the full name of your organization with all details, and you cannot edit the form after submission.
The $20 million condition does not apply to a company that deploys the model for its employees or embeds it into its own product features. Integrators and cloud providers that want to sell customers access to the model itself should check it.
What hardware you need to run it
Hardware requirements are determined by the need to fit all 816 billion parameters into memory. The configurations below are taken from the Hugging Face breakdown. Total GPU memory was calculated by us.
| Option | Hardware | Context | Total video memory |
|---|---|---|---|
| BF16, Transformers | 4 nodes with 8 H200s each, tensor parallelism 32 | not specified | 32 × 141 GB = 4,512 GB |
| NVFP4, Inference Endpoints | 8 × H200 at 141 GB each | 1,048,576 tokens | 1,128 GB |
| NVFP4, Inference Endpoints | 8 × RTX Pro 6000 at 98 GB each | 131,072 tokens | 784 GB |
| NVFP4, vLLM | Blackwell accelerators for which the NVFP4 format is designed, TP = 4 | 131,072 tokens in the example | depends on the accelerator model |
| NVFP4, SGLang | 4 × GB300 (reference configuration) | not specified | depends on the configuration |
| 2-bit GGUF, llama.cpp | Mac with M3 Ultra and 512 GB of memory | not specified | 304 GB in use, 18 tokens/s |
| 3-bit MLX | Mac with M3 Ultra and 512 GB of memory | not specified | about 344 GB, about 21 tokens/s generation |
Three takeaways from this table:
- Long context is expensive. According to Hugging Face, the KV cache for a full 1 million-token context in BF16 takes about 266 GB. If your tasks only need 128,000 tokens, the memory requirements are noticeably lower.
- Compression has a major impact on results. Meta describes NVFP4 losses as negligible. There are no such claims for the 2-bit and 3-bit versions: you need to verify their quality yourself.
- A Mac is good for experimentation, not for production service. 18–21 tokens per second for one user is a normal speed for testing. For a team of dozens of people and agentic scenarios where the model makes hundreds of calls, that is not enough.
Fine-tuning also requires a cluster. In the Hugging Face example, LoRA fine-tuning (rank 16, only attention projections were trained, experts were frozen) took about two hours on 8 nodes with 8 H100s each, or 64 accelerators.
How strong the model is according to Meta
The figures below are taken from the table on the model card. These are results published by Meta; we did not find a full independent comparison as of October 20, 2026. We selected benchmarks that are close to real work tasks: general index, tool use, terminal work, and development.
| Benchmark | Muse Spark Open | Mistral Large 4 | Qwen3.8-2.4T-A95B | GLM 5.3 | Kimi K3 |
|---|---|---|---|---|---|
| Artificial Analysis Intelligence Index | 39 | 38 | 40 | 45 | 44 |
| MCP-Atlas (tools via MCP) | 89.2 | — | 84.5 | 84.2 | 82.3 |
| Terminal-Bench 2.1 | 76.4 | — | 82.0 | 83.9 | 85.0 |
| DeepSWE 1.1 (development) | 60 | 62 | 57 | 67 | 68 |
| OSWorld Verified (OS interface work) | 80.4 | — | — | — | 84.8 |
“—” means Meta did not provide a result. Based on Meta’s own table, the model is strong at tool use via MCP, but trails GLM 5.3 and Kimi K3 in terminal and engineering tests and on the overall index. So the new release alone is not a reason to abandon an open model you have already chosen: first, test it on your own tasks. We covered how to build such a set in the article “Your Own Benchmark for Choosing an LLM”.
Limitations for a Russian company
In addition to hardware, a Russian company has four questions that need to be answered before a pilot.
- Meta status. On March 21, 2022, Moscow’s Tverskoy District Court designated Meta Platforms as an extremist organization and banned its activities in Russia, reports Pravo.ru. The same report says that Russians’ use of blocked social networks is not considered extremist activity. The court did not comment on the use of Meta models in business. Assess how this may affect your company together with a lawyer, especially if the company works with government contractors.
- Access to the weights. Access is granted manually upon request with the organization’s details, and the license requires compliance with trade restriction laws. Whether a Russian organization’s request will be approved is unknown: we did not submit one. Downloading weights from third-party mirrors around moderation is a risk both to the license and to file integrity.
- Hardware. The configurations in the table above are built on NVIDIA H200, B200, and RTX Pro 6000 accelerators. For critical information infrastructure facilities, there are separate restrictions on foreign hardware—we covered them in the article “Foreign GPUs for AI at CII Facilities”.
- Cloud access. According to TechCrunch, the model immediately became available through several API platforms, and these are overseas services. Sending personal data there means cross-border data transfer, which triggers the requirements of Federal Law 152. Your own servers remove that issue, but bring back the hardware question.
Who should consider the model, and who shouldn’t
| Situation | What to do |
|---|---|
| The company already has a cluster on H200 or Blackwell and internal agent workflows for development | Add Muse Spark Open to the comparison with your current model on your own test set |
| You need an on-prem LLM, but the hardware budget has not yet been approved | Compare total cost with smaller models; this model sets the upper limit on requirements |
| You need a Russian-language assistant for employees without complex agent chains | First check quality in Russian; if results are equal, the smaller model will be cheaper to operate |
| The company works with government contractors or falls under CII requirements | Get a written opinion from legal and information security before any experiments |
| You need a quick experiment without your own hardware | Test only on anonymized data, taking Federal Law 152 into account |
If you are still choosing between cloud and your own server, start with the calculation in the article “Cloud or On-Prem LLM: TCO Calculation”. The model and hardware review for 2026 is in the article “On-Prem LLMs 2026”.
What your company should do now
- Do not buy hardware because of a news headline. In a few weeks, independent evaluations and early implementation experience will appear, and the picture will become clearer.
- List the tasks that require your own model. For example: an agent for development in a closed environment, search across internal documents, or handling incoming emails. For each task, specify whether it needs long context and image processing.
- Update the test set. Take 50–100 real prompts with correct answers and evaluation criteria and add Muse Spark Open to the comparison alongside your current model.
- Calculate memory for your context. Decide how many context tokens you actually need: that determines whether eight accelerators with 98 GB each are enough or whether you need 141 GB each.
- Send three documents to your lawyer: the Muse Model License, the usage policy, and the notice about Meta’s status in Russia. Ask for a written conclusion specifically for your scenario.
- If you decide to try it, submit a request for official access on behalf of the organization and record the response. A denial is also a useful result: it removes the model from the list without hardware costs.
For agents that work with code, it helps to think ahead about isolation and access rights—details in the article “AI Agent for Development in a Closed Environment”.
How we verified the information
- Sources: the model card and technical analysis on Hugging Face, the TechCrunch article from October 15, 2026, and the Pravo.ru report from March 21, 2022. We checked the repository metadata (last modified on October 15, 2026, manual access moderation) through the Hugging Face API on October 20, 2026.
- Test selection: from Meta’s table, we took five tests related to tools, terminal, development, and interface work, and five models for which the table provides data.
- Calculations: total video memory is the number of accelerators multiplied by the memory size of one accelerator, based on Hugging Face data. There were no other calculations.
- What we did not do: we did not run the model, did not measure Russian-language quality, and did not submit an access request for the weights. Cloud platform prices are not included in the article—we could not verify them from primary sources.
Frequently asked questions
Can Muse Spark Open run on a single server?
Yes, if the server has eight accelerators with a large amount of memory. According to Hugging Face, the NVFP4 version runs on eight H200 cards with 141 GB each with a 1 million token context, or on eight RTX Pro 6000 cards with 98 GB each with a 131,072-token context. On a standard server with one or two GPUs, the model will not fit, and the 2-bit version for Mac is suitable only for experiments.
Can the model be used in a commercial product?
Yes, the Muse Model License allows commercial use and fine-tuning. You must comply with the usage policy and prominently display “Built with Muse Spark Open.” If you sell access to the model to third parties as a service, and your company’s revenue with affiliates exceeds $20 million over 12 months, you need a separate agreement with Meta. For a Russian company, also assess Meta’s status with a lawyer.
How does Muse Spark Open work with Russian?
There is no exact data. The model card says it was trained on data in more than 100 languages, but not all of them were tested, and quality may drop in less-supported languages. Meta has not published results in Russian. The only reliable way is to run the model on your own Russian-language prompts alongside your current model and compare the answers using the same criteria.
How is Muse Spark Open different from Muse Glimmer?
Muse Spark Open is a large model in the Muse series, while Muse Glimmer, which Meta released in August, is a small dense model from the same series for local deployment. They share a tokenizer and vocabulary. For business use, the difference is in requirements: Glimmer is designed for modest hardware, while Muse Spark Open requires a cluster and has undergone additional safety training to operate without system restrictions.
Should you wait for independent tests?
Yes, if there is no urgent need. As of October 20, 2026, the quality figures in the model card were published by Meta itself, and we did not find a full independent comparison. Until one appears, prepare a test set and a legal opinion: then, when independent data becomes available, the model can be tested within a few days.
Primary sources
- Hugging Face: Muse Spark Open model card
- Hugging Face: Welcome Muse Spark Open — technical analysis
- TechCrunch: Meta launches an open-weight AI model that runs on Macs, 15.10.2026
- Pravo.ru: court decision recognizing Meta as an extremist organization, 21.03.2022
How AIrassvet helps evaluate an open model for your own servers
A new open-weight model is not a ready-made solution, but a candidate that needs to be tested on your data and your hardware. On the AIrassvet website The areas relevant to this evaluation are listed below:
- Process audit: which tasks really require their own model and what context they need;
- Deploying and fine-tuning LLMs on the customer’s own servers, including inference optimization for the available hardware;
- Testing candidates on a set of real queries before deciding to buy hardware;
- Integrating the selected model into AI agents and business systems.
The first step is to choose one process and define its current state, data sources, constraints, and the criterion you will use to determine whether the model is a fit.
Bottom line: a release for the test set, not for procurement
Muse Spark Open is a serious open-weights model, but it does not change the equation for most companies. To run it in-house, you need eight high-memory accelerators, a legal review of the license and Meta’s status, and quality testing in Russian. So far, only Meta itself has published the quality metrics.
Use the release as a prompt to get organized: build a test set, calculate the required context, and send the license to legal. Then the next strong open-weights model — from Meta or any other developer — will require a few days of review, not a new project.
