Qwen3.8-Flash-Next 177B Hits 11-15 tok/s on a Single RTX 5070

On October 3, 2026, a post on the LocalLLaMA subreddit described a local inference setup for Qwen3.8-Flash-Next 177B. According to the author, he had been working on a llama.cpp-based expert streaming approach on Windows, running the model at UD-IQ3_XXS quantization on a single RTX 5070 with 12GB of VRAM, alongside 32GB of DDR4-2400 memory, a Ryzen 5 5600GT, and a PCIe Gen3 link. The headline figure for the setup is a range of 11-15 tok/s on that one GPU.

The post reports several numbers behind that range. The benchmark came in at about 11.5 tok/s, which the author says is up from roughly 7 tok/s on the setup he inherited. In ordinary conversation he saw 14-15 tok/s. On a long coding prompt, the model generated 4,892 tokens at 10.15 tok/s and produced a working single-file Snake game.

The post does not list context length, batch size, or other generation parameters, and the point of comparison is the author's own inherited setup rather than any published baseline. So far this is the only report on the configuration: the figures come from the poster's own account, and the material contains no third-party reproduction or independent verification.