Ciru Inference Lab Strix Halo research report

ROCmFPX Quality Lane Comparison

Ultra Quality 7.61 BPW: What Changed From the 7.37 BPW Quality Model

A focused comparison of the published Qwable ROCmFP6 Quality report and the newer Qwable 3.6 27B Chadrock ROCmFPX UltraQuality lane. UltraQuality spends roughly a quarter extra bit per weight, replaces the older STRIX QUALITY recipe as the default quality build, and recovers MTP draft behavior while reducing KLD drift.

Ultra BPW 7.61 7.6146 BPW, 26,004,616,416 bytes
Quality BPW 7.37 published Strix Quality quant report
BPW Spend +0.24 about +3.3% over Quality
Ultra Refresh TG 25.92 20KB served MTP prompt, 512 generated tokens

Result Summary

The earlier Quality model is the public 7.37 BPW Strix Quality lane from the June ROCmFP6 report. It already matched the Unsloth Q6 baseline on HermesAgent-20 plus-pass count and slightly beat it on aggregate score in that run.

The UltraQuality lane is the newer 7.61 BPW default quality build: ranked leave-32 ROCmFPX with a Q6_K splice and a Strix ROCm MTP profile. It is not just a label change. The refresh evidence shows better MTP acceptance, lower KLD, and a higher HermesAgent-20 score than the superseded STRIX QUALITY row.

Interpretation: Ultra Quality is the quality-first lane. Use it when the goal is reducing tail risk and preserving high-impact attention behavior. Use 7.37 BPW Quality when the goal is the best published balance of quality evidence, size, and served speed.

Qwable 3.6 27B ROCmFPX ranked attention leave-32 Strix mmid5

Quant Policy Delta

Policy Table

Lane BPW Quant size Policy
ROCmFP6 Quality 7.37 24018.32 MiB Q6_0_ROCMFPX quality mix with 312 Q6 tensors and 194 Q8 tensors.
ROCmFPX Ultra Quality 7.61 26,004,616,416 bytes Ranked leave-32 ROCmFPX recipe with Q6_K splice, promoted under the UltraQuality public name.

Quality Evidence

The older Quality lane has the June public report evidence. UltraQuality has a fresh card-refresh evidence set: served MTP speed, PPL/KLD, HermesAgent-20, and EvalPlus.

Metric 7.37 BPW Quality 7.61 BPW Ultra Quality
HermesAgent-20 0.78 score 0.82 score
HumanEval+ 155/164 plus passes 154/164 plus passes, 160/164 base passes
PPL/KLD lane Refresh PPL 6.5097 +/- 0.0928; mean KLD 0.007113; P99 KLD 0.057581; same-top 96.495% Refresh PPL 6.5212 +/- 0.0932; mean KLD 0.002420; P99 KLD 0.019161; same-top 97.843%
What changed Broad Q6/Q8 promotion, good enough to prove the direction More surgical ranked restore recipe with much better draft acceptance and lower KLD

Served Speed

Same-Prompt Refresh Rows

Lane Prompt Generated TG tok/s PP tok/s Draft accepted
7.37 BPW STRIX QUALITY 3946 512 20.387 217.333 387/733
7.61 BPW UltraQuality 3946 512 25.921 209.835 437/439

Speed rows are served-MTP measurements using the same 20KB prompt shape and 512 generated tokens. The earlier 8.37 tok/s STRIX QUALITY refresh row is excluded here because it was superseded by the fixed June 30 real-served verification.

Evidence Boundary

Confirmed for Ultra: BPW, file size, rankleave32/Q6_K-splice recipe, Strix ROCm MTP profile, served speed, PPL/KLD, HermesAgent-20, and HumanEval+.

Confirmed for Quality: published BPW/size, tensor mix, HermesAgent-20 aggregate, HumanEval+ result, PPL, and historical speed rows from the earlier report.

Comparison boundary: the headline speed comparison uses the fixed same-prompt served evidence. The June 24 public report remains the source for the older report's original matrix and provenance.

Sources

This page uses the published ROCmFP6 Quality report data, the UltraQuality model card evidence, the quantization inventory, the fixed served-speed stores, PPL/KLD logs, HermesAgent-20, and EvalPlus rows from the lab. Internal paths and filenames are intentionally omitted from the public report body.

Open the original Quality report

Generated July 1, 2026. Public labels use the requested 7.61 BPW suffix.