Jackrong-Qwopus3.6-27B-Coder EXL3 4.5bpw h6
This repository is a direct EXL3 quantization of the public model
Jackrong/Qwopus3.6-27B-Coder.
It is not an Anderson Operator fine-tune and it was not trained by this build.
Expected repository names:
Jackrong-Qwopus3.6-27B-Coder-EXL3-4.5bpw-h6-textonly-nonmtpJackrong-Qwopus3.6-27B-Coder-EXL3-4.5bpw-h6-vision-nonmtp
The textonly-nonmtp variant strips both vision and MTP tensors/config for a
text-generation fallback. The vision-nonmtp variant preserves vision
tensors/config and strips MTP only.
Quantization
- Source model:
Jackrong/Qwopus3.6-27B-Coder
- Source repo SHA recorded during build:
f22041be82998a006c9fde642c7900fb42aff94e
- Quantization: EXL3
- Target bitrate: 4.5 bpw
- Head bits: 6
- Calibration: custom agent calibration parquet used by the Anderson Operator quants
- MTP: stripped
- Vision: depends on repo name
Positioning
This artifact exists as an apples-to-apples serving-format fallback for users who
want the public Jackrong coder model in the same EXL3 4.5bpw/head-bits-6 format
as the Anderson Operator variants.
Do not cite this repository as a new training run. Training details, datasets,
and benchmark claims belong to the upstream Jackrong model card unless rerun and
verified for this exact quantized artifact.
Evaluation
No new SWE-bench result is claimed by this quantization card. If benchmark
results are added later, record the exact artifact, harness version, denominator,
serving backend, prompt/template, and raw scoring output.
Caveats
- This is a quantized derivative of the upstream public Jackrong model.
- Check upstream licensing and use restrictions before redistribution or
commercial use.
- Vision serving support depends on the client stack; use the
textonly-nonmtp
variant when a text-only ExLlamaV3/TextGen path is desired.
