A Better INT4 Qwopus We Decided Not to Ship
An improved INT4 build of a 27B coder model got closer to full precision on every quality measure — and still did not ship, because it was slower and broke the fast drafter most people use.
Summary of a report from Launch80, the NSS research lab. Read the full report →
Launch80 went back to their INT4 build of Qwopus3.8-27B-Coder-Flash to close the quality gap to the best 4-bit builds. The new checkpoint came out closer to full precision on every next-token measure. They still kept the previous release.
What changed
- An AutoRound asymmetric quantizer for the model body, which beat GPTQ on the same calibration text.
- A finer-grained output head (group size 32 instead of 128), which recovered most of the top-1 agreement the INT4 head had cost.
- Mean KL divergence against BF16 fell from 0.0412 to 0.0327, and top-1 agreement rose from 92.84% to 93.66% — clearing a bar v1.0 had missed. No task lost significantly against BF16.
Why it didn't ship
Speed. Decode with the built-in MTP head dropped from 75.8 to 72.5 tok/s, and vLLM refused to start the INT4 DFlash2 drafter that gives v1.1.1 its 92.5–96.0 tok/s headline. A release that is better on paper but drops the profile most people actually run is not an upgrade, so v1.1.1 stays.
The report is candid about process, too: the candidate sat on the model repo's main branch for about four minutes before the drafter result was weighed. It was reverted and never tagged, and the rule is now that every gating result is read before anything is published.
Read the full report on Launch80
Methodology, every table and chart, and the caveats this summary leaves out.
launch80.com/blog/qwopus-coder-int4-v2-study →