Auto-quantized from upstage/solar-pro-preview-instruct by fastserve's self-quantization pipeline (publish/) — so it can be trusted the way a random community AWQ requant sometimes can't. Two community quants fastserve auto-detected during its own benchmark run looped garbage tokens instead of answering; every checkpoint published here passed an accuracy gate against its bf16 baseline first.

Within 0.1 (absolute) of the bf16 baseline, <30% degenerate (repeated-token loops). A checkpoint that failed this gate would not have been uploaded.

fastserve auto-detects this checkpoint — point it at the original model id and it finds this AWQ + wires up speculative decoding:

Then query it like any OpenAI endpoint:

Inherits the base model's license — see upstage/solar-pro-preview-instruct for terms.