Meta and NVIDIA widen the choice of downloadable AI models
Meta’s Muse Spark Open targets substantial infrastructure, while NVIDIA’s new Bolt model targets a 24 GB memory budget: downloadable weights now span very different deployment choices.
Meta released Muse Spark Open on October 15 with 816 billion parameters, of which 42 billion are active for each token.
That distinction reduces the work performed at each step; it does not remove the need to store the other parameters. Meta’s model card specifies at least 700 GB of GPU memory for its NVFP4 version and 2 TB for BF16. This is infrastructure-scale software. [1][2]
NVIDIA’s same-day release takes a different direction. Nemotron 3.5 Bolt’s 24 GB variant compresses a larger model to fit a much smaller memory budget; its research report describes a 74-billion-parameter model with nine billion active parameters. The announced 4 GB variant is scheduled for October 23, so it is not an available option today. [3][4]
For a team considering self-hosting, the useful comparison starts with its hardware, workload and permitted use. Neither downloadability nor a parameter count establishes comparable quality. Artificial Analysis’s separate measurements found that Bolt traded some performance for compactness relative to NVIDIA’s larger Super model. Meta’s weights also require acceptance of its custom licence. Test the actual quantization and workload before committing infrastructure. [2][5]
Take it further
Inspect Muse Spark Open requirements Public model card; downloading requires accepting Meta’s terms. Match the model format and memory requirement to your infrastructure.
Inspect Bolt’s deployment assumptions Public technical report. The 24 GB target is configuration-specific; the smaller 4 GB release is scheduled for October 23.
Sources & details 5 sources
Discuss in your voice app
Take the reporting, visual explanations and original sources into a conversation. Copy the brief, paste it into your preferred assistant, then start its voice mode.
Paste the brief into your voice app to discuss this reporting.Reporting sources
- Meta · Muse Spark Open release
October 15 announcement; architecture and availability are vendor descriptions.
- Meta · Model card and licence
Original model repository. Specifies 700 GB NVFP4 / 2 TB BF16 memory floors and gated acceptance of custom terms.
- NVIDIA · Super and Bolt availability
October 15 announcement; separates the released Bolt variant from the October 23 scheduled smaller model.
- NVIDIA Research · Bolt 24 GB
Original technical description of compression and the intended memory budget; not a guarantee for every inference configuration.
- Artificial Analysis · Independent measurements
Original October 15 measurements of NVIDIA models. Not a head-to-head evaluation against the newly released Muse Spark Open.
Images & figures
AI-generated editorial illustration of sparse activation within a model whose complete weights remain stored; not documentary evidence or a measured diagram.
Euangelion · AI-generated editorial illustration · Source · Usage terms