The 2.0 bpw EXL3 quant of Qwen3.8-27B occupies 9.7 GB on disk and is the minimum the serving kit will load onto a 16 GB NVIDIA card. The kit, from MiaAI-Lab, selects a quant at setup based on detected VRAM and exposes an OpenAI-compatible endpoint. Everything at 2.5 bpw and above is pulled from turboderp's repository by revision; the 2.0 bpw file is the one artifact that lives outside that upstream.

Who is constrained is the buyer, not the model. A 27B-parameter model at 2.0 bits per weight is heavily compressed, and the document does not measure what that compression costs in output quality or in tokens-per-second at any given context length. The installer proves the weights fit in 16 GB. It does not prove the model is useful at that quant level. For a datacenter buyer, the relevant number is that 16 GB is the floor, not the target: the document lists disk sizes up to 22.9 GB without stating which card that maps to.

No latency figure, no tokens-per-second, no context-window ceiling, and no quality comparison between the 2.0 bpw and 2.5 bpw quants appear in the repository. The document is an install script with a quant-selection table. It tells you the weights will load. It says nothing about what the model will produce.

The kit proves the weights fit in 16 GB; it does not prove the model is useful at 2.0 bpw. The document leaves out every performance number a buyer would need to compare against a 2.5 bpw or higher quant.

No inference latency, tokens-per-second, context-window limit, or quality comparison between quant levels appears in the repository.

The 2.0 bpw EXL3 quant is the floor for a 16 GB card. No latency, throughput, or quality metric is provided. The buyer cannot determine whether the 27B model at this quant level is acceptable for production use from this document alone.