AI Briefing
KO

27B won

·2026.04.23 12:52

Key point

On MacBook Pro M5 Max, 35B-A3B was faster, but 27B was cleaner.

Details

Compared two models on a MacBook Pro M5 Max 64GB using Google TurboQuant.

When asked to draw a waveform in HTML, Qwen3.6 35B-A3B responded quickly at 6672 tokens / 2m 10s / 65 tok/s, but the result was evaluated as rough and less polished.

On the other hand, Qwen3.6 27B took longer at 7344 tokens / 5m 22s / 24 tok/s, but produced cleaner and more consistent output.

Based on this, the author summarized:

  • 27B is better suited for tasks where structure and planning matter
  • 35B-A3B is advantageous when simple, fast responses are needed

The inference server mentioned was atomic.chat, with related code in the Atomic-Chat repository.

qwenquantizationllama.cpphtml
Read originalShare on Threads

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

RedditTextCLF Releases Calibration-Free TQ Quantization Method and Open-Source Quant FactoryRedditUkisAI Releases Swift Series: Efficient Qwen-Based LLMs with Reduced Thinking TokensPyTorchKRPrismML Releases 'Ternary Bonsai 2 27B', a Ternary Quantized Model Maintaining 98.2% of FP16 PerformanceHacker NewsApple Releases LensVLM-9B for Processing Compressed Text ImagesTLDR AIQwen Releases 'RecreationWorld' for Training AI Agents to Rebuild AppsTLDR AIQwen Intelligence Launches Three Mobile AI Agents and Open Benchmark Suite