AI BriefingQwen3.6-27B Merge, ARC 0.699
Key point
A NuSLERP merge based on Qwen3.6-27B achieved 0.699 on ARC Challenge.
Details
A NuSLERP merge based on Qwen3.6-27B has been released, recording ARC 0.699 on mxfp8.
The base is nightmedia/Qwen3.6-27B-Architect-DS9 and DavidAU/Qwen3.6-27B-Heretic2-Uncensored-Finetune-Thinking, and it was stated that Claude 4.6, Polaris Alpha, Star Trek TNG, and Philip K Dick distills were mixed in.
The key benchmarks are as follows.
mxfp8: arc 0.699, arc/e 0.871, boolq 0.910q8-hi: arc 0.694, arc/e 0.865, boolq 0.910mxfp4: arc 0.694, arc/e 0.872, boolq 0.909
Runtime metrics by quantization were also released.
q8-hi: perplexity 3.895 ± 0.025, peak memory 37.26 GB, 215 tok/smxfp8: perplexity 3.921 ± 0.025, peak memory 34.74 GB, 218 tok/sqx86-hi: perplexity 3.898 ± 0.025, peak memory 32.36 GB, 218 tok/sqx64-hi: perplexity 3.918 ± 0.025, peak memory 25.64 GB, 217 tok/smxfp4: perplexity 3.999 ± 0.025, peak memory 21.30 GB, 225 tok/s
The baseline model Qwen3.6-27B-Instruct's mxfp8 results were arc 0.647, arc/e 0.803, boolq 0.910. This merge raises ARC-series scores while showing memory usage in the 21.30GB~37.26GB range and throughput of 215~225 tok/s depending on quantization.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.