Meta has released Muse Spark Open, an open-weight model with 816 billion parameters and a one-million-token context window. The company calls it the most capable open model in the US and aims it at coding and agent-style tasks. It is designed to run locally on Macs, but needs at least 1 TB of memory, ideally on a Mac Studio with the M5 Ultra chip. It is already available through several API platforms.

What it means: An open model of this size changes the maths for anyone building products on AI. Until now the choice was usually between expensive closed models behind an API and smaller open models with weaker results. If independent tests confirm Meta's benchmarks, businesses gain a strong option they can host themselves. For a Greek software house or bank, that means sensitive material such as source code, contracts or customer records does not have to leave for an outside provider, which makes GDPR and DORA compliance easier to argue. The 1 TB memory requirement, however, shows that running it locally is not for every office. A Mac Studio with that much memory costs many thousands of euros, so for most small and mid-sized firms the realistic path is hosting on a European cloud or using it through an API with clear data-processing terms. Before adopting it, teams should read the licence carefully, because open weights do not always mean unrestricted commercial use. The broader point is that competition among open models pushes prices down and gives smaller teams more bargaining power with the large providers. Procurement teams should benchmark it against their current model on their own tasks before switching anything in production.

Read the original on TechCrunch

Related service: Custom software

Related stories