Fork of FreeToken with DeepSeek-V4.1, vision and speculative decoding (2x3090 numbers inside)

A developer has forked FreeToken, integrating DeepSeek-V4.1, vision capabilities, and speculative decoding, running on a 2x3090 setup. The fork aims to improve performance, with initial tests showing that while CUDA graphs on the plain path are a net loss, a per-batch-size verify graph boosted the MTP path from 16 to 28 tok/s in isolated runs. The developer acknowledges it's a work in progress and seeks feedback on models and quant formats.