The source describes GSQ as learning grid assignments and per-group scales for each tensor through a Gumbel-Softmax relaxation. RCO selects a quantization type for each tensor while keeping the complete model within a size budget. The page also identifies the model as a mixture-of-experts architecture, routing tokens through selected expert feed-forward blocks rather than activating every expert for every token. The supplied source does not provide benchmark results, memory measurements, or deployment comparisons, so it establishes the build and its stated method, not a measured performance or quality advantage.
Raw SignalOct 11, 20261 min read1 source
The upstream ISTA-DASLab build combines GSQ and RCO.
In one minute
ISTA-DASLab has published a GGUF quantized build of Qwen3.8-Flash-Next combining GSQ and RCO.