DeepSeek-V4-Flash-Vision-Exp (285B MoE) on 10-12x RTX 3090 — spec decoding, vision

The DeepSeek-V4-Flash-Vision-Exp (285B MoE) model has been successfully run on 10-12x RTX 3090 GPUs, achieving 60+ tok/s decode on 10 GPUs and 120+ tok/s on 12 GPUs. This setup supports vision, speculative decoding, and tool calls, with a 1M context without offload or 4M with RAM offload. The implementation uses FP4 experts and FP8 attention, with 157 GB weights, and is fully documented and reproducible via a Docker image and GitHub repository.

This is the first public report of DeepSeek-V4-Flash-Vision-Exp running on consumer-grade Ampere GPUs, unlike previous benchmarks on more powerful hardware.