Deepseek V4 Flash Vision Exp Local Deployment

1 update · 1 outlet · 9 Sep 2026

Timeline

1. 9 Sep 2026DeepSeek-V4-Flash-Vision-Exp runs on 10-12 RTX 3090s at 60-120 tok/s

   A reproducible path to running a 285B MoE vision model with speculative decoding at 60-120 tok/s on consumer RTX 3090s could shift what counts as feasible for local, offline deployment of frontier-scale models.

   1 source · thin-sourcing, unconfirmed

All sources (1)

1. r/LocalLLaMA2026-09-09

Related subjects

• Cosmos3 Int4 Quant Local ReleaseLocal AI Scene
• Mentria Browser Inference EngineLocal AI Scene
• Minnow Llada2 2 Inference ServerLocal AI Scene