NVIDIA's Nemotron 3.5 Bolt 74B MoE fits 24GB via 2-bit quantization
CulturalKing5623 · reddit · 2026-10-14
- NVIDIA will release Nemotron Bolt 3.5 74B on Oct. 15: a 74B-parameter MoE with 9B active parameters, packed into 24GB using 2-bit quantization with quantization-aware training to retain quality.
- Positioned as a speed/intelligence balance: faster than most dense models in its memory class, smarter than most popular MoEs with similar footprints.
- A smaller Nemotron 3.5 Bolt E4B 4GB version lands Oct. 23.
- Aimed at local hardware, with imagery showing NVIDIA Local AI running models on laptops.
More from Infra
- Perplexity's SPACE sandbox runs millions of daily executions with sub-0.0002% error rate — inductionheads · 2026-10-15
- SkyPilot v0.14.0 ships dynamic job groups, multi-user Slurm GA and AWS Trainium support — skypilot_org · 2026-10-15
- OpenAI puts compute CTO Uday Ruddarraju in charge of infrastructure as Sachin Katti exits — rohanpaul_ai · 2026-10-14
- Running Qwen 3.8 flash with 256K context on a 12GB GPU: open-source local setup hits 30-38 tok/s — Present_Resource4022 · 2026-10-14
- CoreWeave bonds blow out to 15.1% yields as distress spreads across the curve — SumitGup · 2026-10-14
- Ben Bajarin: clean rooms and data center footprint are the two big drivers of the AI cycle — BenBajarin · 2026-10-14