Meta releases 800-billion-parameter open model requiring 1TB of memory
Meta has released Muse Spark Open, an 800-billion-parameter open-weight AI model designed to run on Apple Silicon. The release extends Meta’s push for open models and matters because the company says the model leads US open-weight systems, while its benchmarks show advantages over two prominent Chinese models on selected tasks.
Running it locally requires about 1TB of unified memory, which Meta says can be provided by linking two M5 Ultra Mac Studios, each with up to 512GB. Alternatively, four 256GB machines would cost roughly $38,000; Apple has not yet given a price or exact shipping date for the 512GB configuration, expected in late October. The model’s weights are available to download from Hugging Face, though its reported performance is based on Meta’s benchmarks.
- Meta released its 800-billion-parameter Muse Spark Open model.
- Local use needs about 1TB of unified memory.
- Four 256GB Mac Studios would cost roughly $38,000.
New here? Start with this
An AI model is a computer system trained to produce responses or other outputs from patterns in data. Its “parameters” are the internal values it adjusts during training; an 800-billion-parameter model is very large, though that figure alone does not show how well it performs.
Meta is a US technology company that develops AI models, including models whose weights—the values learned during training—are available for others to download and use. Such models are called open-weight models, but access to the weights does not by itself make a system easy or affordable to run.
The memory requirement matters because it affects who can use the model locally and what equipment they need. Comparisons with other US and Chinese models may offer a point of reference, but benchmark results depend on the tasks and tests chosen.
Both sides, in good faith
The strongest fair case each way — we don't pick a winner.
The case for
Releasing the weights gives researchers and developers more access to a powerful model, supporting experimentation, transparency and competition beyond a small number of closed AI providers. Meta’s reported benchmark results suggest it may also strengthen the position of US open-weight models, while running it on linked Apple hardware offers one route to local use.
The case against
Calling the model open may overstate how accessible it is in practice: its roughly 1TB memory requirement puts local deployment beyond the reach of most individuals and organisations. Its performance claims also rely on Meta’s own benchmarks, so independent evaluation is needed before comparisons with other models can be treated as conclusive.
AI Americas Business Companies Geopolitics Politics Technology World
Read the full article at the source →
Originally published by Engadget as “Meta built a huge open-weight AI model for Macs, but you’ll need 1TB of RAM to run it”.