Meta Muse Spark Open: local AI in the office and what it means

What happened

On 15 October 2026, Meta released Muse Spark Open, an open-weight AI model with 816 billion parameters, which Meta itself calls the most capable American open model to date. The key point for businesses is that, as Apple and Meta state, this model can be run locally on two desktop computers in an office, rather than only in the cloud of a large provider.

What this means for companies in Latvia

First, the model itself. According to Meta's model card, it is a mixture-of-experts model in which 42 billion parameters are active for each token. According to the same model card, the model can process up to 1 million tokens and accepts text, images, video and audio, but responds in text. Meta states that the model is intended for programming, AI agent work, long tasks and multi-step reasoning.

Now the hardware. Apple and Meta state that the model can run on two Apple M5 Ultra Mac Studio computers, each with up to 512 GB of unified memory (around 1 TB in total), using Apple's MLX framework. According to Apple and Meta, four computers with 256 GB of memory each also work. Apple and Meta claim that running the model locally without the cloud provides an additional layer of security and can reduce costs. According to Apple and Meta, a Mac Studio with M5 Ultra and 256 GB of memory costs from 10,799 USD, and, as Meta states, the licence permits limited commercial use.

The limitations matter too. As VentureBeat writes, the results published by Meta show a model with specific strengths rather than an overall leader. VentureBeat notes that, according to the model card, Muse Spark Open scores 39 on the Artificial Analysis Intelligence Index, while the independently published result for MiMo-V2.6-Pro is 46. As VentureBeat writes, in terminal programming and long software tasks the model lags behind Kimi K3 and GLM 5.3. VentureBeat also stresses that the announcement does not prove a cost advantage, as it gives no prices per token in the cloud.

Our view (AI-Lab Space opinion): the trend is clear. Powerful AI no longer has to run only in the cloud of a large provider; it can also be kept in your own office. For companies with sensitive data this opens up new possibilities:

• Finance and accounting - invoices, contracts and client financial data can stay inside the company.
• Healthcare - patient information can be processed without sending it to external service providers.
• Legal services - contract analysis and document preparation without handing confidential texts to third parties.

In such cases, keeping data in-house can make GDPR compliance considerably easier, because it is clearer where the data is and who has access to it.

But we must be honest: not every company needs such a large model and such expensive hardware. In our experience, a smaller local model, or a model run in an EU cloud, is often more cost-effective and entirely sufficient for everyday tasks.

How to use it in practice

Consider a typical situation. In a company where contracts and requests arrive by email, employees read them manually, extract the key information and enter it into a system. Before introducing AI, the question often arises: are we allowed to send these documents to an external AI service?

Before: documents are processed by hand because management does not trust cloud solutions or client contracts do not allow it. The process is slow, and errors appear during retyping.

After: the AI model runs on the company's internal infrastructure or in a chosen EU cloud. It reads the document, extracts the main fields, prepares a summary and passes the data to the CRM or the company's system. A person checks the result and approves it. The data never leaves the controlled environment.

To make such a solution work in the long term, we recommend following a few principles:

• Start with the task, not the model. First define exactly what the AI should do: document processing, answering customer questions, internal knowledge, programming assistance.
• Check the licence. Open weights do not mean complete freedom. If the licence permits only limited commercial use, you need to understand whether your case falls within it.
• Compare total costs. Hardware is only one part. Add maintenance, updates, electricity, staff time and a backup plan in case a computer fails. Then compare this with cloud costs at your actual level of usage.
• Build so that the model can be replaced. AI models change very quickly. If a solution is tied to one specific model, you may have to rebuild everything in six months. It is better to build a layer between the business process and the model, so that the model can be swapped for another without major rework.
• Test with your own data. Public results are a guide, but what matters is how the model copes with your company's documents and language, including Latvian.

Where to start

If you are thinking about local AI, we do not recommend starting by buying hardware. A better path looks like this:

1. Choose one process that involves sensitive data and a lot of manual work, for example contract or invoice processing.
2. Assess data sensitivity and the requirements set by GDPR and your contracts with clients.
3. Try a smaller model or a solution run in an EU cloud to see whether the quality is sufficient.
4. Calculate total costs for both options: local and cloud.
5. Only then decide on hardware, if the trial shows that a larger local model really delivers a significant benefit.

Muse Spark Open is a good signal that powerful AI is becoming more accessible beyond the clouds of large providers. But the right choice will differ for every company.

If you want to understand whether local AI suits your company, contact AI-Lab Space. Together we will assess your processes, data requirements and costs, and recommend a solution that fits your needs rather than the latest news.