Project Maya Runs 321B-Parameter GLM-5.3-Flash on One GPU

Open-source Project Maya runs the 321B-parameter GLM-5.3-Flash locally on one NVIDIA GPU by tiering experts across GPU, RAM and SSD, reaching 30+ tokens/s and up to 118 tokens/s on four RTX 4090s while retaining 97.7-99.2% of FP8 zero-shot accuracy.

• Nvidia confirms RTX 5090 not discontinued, will keep shipping2026-10-17
• NVIDIA Reportedly Discontinuing RTX 5090, GB202 Reserved for RTX PRO2026-10-10
• RTX 5090 purchases reportedly require no-export paperwork2026-10-02
• Radeon RX 10800 XT may beat RTX 5090 by 15-25% in 4K gaming and local AI2026-09-20
• RTX 5090 Vanishes from US Retail2026-09-14