AI BriefingThree Models Merged
Key point
A merged model combining Qwen3.6-35B-A3B with Qwopus and Holo3, along with benchmarks, has been released.
Details
Qwen3.6-35B-A3B-Holo3-Qwopus, which merges Qwen/Qwen3.6-35B-A3B, samuelcardillo/Qwopus-MoE-35B-A3B, and Hcompany/Holo3-35B-A3B, has been released.
The distribution is provided in bf16, qx64-hi-mlx, and mxfp4-mlx, along with performance figures for each inference/quantization variant.
- bf16: perplexity 4.217 ± 0.027, peak memory 76.15 GB, 1642 tok/s
- qx64-hi: perplexity 4.231 ± 0.028, peak memory 36.83 GB, 1573 tok/s
- mxfp4: perplexity 4.522 ± 0.030, peak memory 25.33 GB, 1609 tok/s
In Instruct mode, measured on mxfp8, results were arc 0.608, arc/e 0.770, boolq 0.897, hswag 0.761, among others. Component metrics compared with other variants in the same family were also released, showing that Holo3 Instruct scored generally higher than both Qwopus Instruct and Qwen Instruct.
In Thinking mode, qx64-hi recorded arc/e 0.476, boolq 0.708, hswag 0.693, while qx64 was reported with peak memory 30.69 GB, 1366 tok/s, and perplexity 4.702 ± 0.032.
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.
Related
Apple Releases LensVLM-9B for Processing Compressed Text Images
Qwen Releases 'RecreationWorld' for Training AI Agents to Rebuild AppsStep 5 Preview Released: 600B Parameter MoE Model Highlights Agentic Tasks and Financial Analysis Performance