AI Briefing
KO

Qwen3.6-35B tops the charts

·2026.04.22 20:22

Key point

The little-coder and Qwen3.6-35B-A3B combo scored 78.7% on Aider Polyglot.

Details

Running Qwen3.6-35B-A3B through the little-coder + llama.cpp combo on the Aider Polyglot 225-question set produced 177/225 = 78.67%, placing it in the top-10 band of the public leaderboard.

  • On the same hardware and the same harness, two previous runs of Qwen3.5 9B scored 104/225 = 46.22% and 101/225 = 44.89%.
  • The test environment was RTX 5070 Laptop 8GB VRAM, Intel i9-14900HX, 32GB DDR5, with llama.cpp built for CUDA 13.1 / sm_120.
  • The model was unsloth/Qwen3.6-35B-A3B-GGUF UD-Q4_K_M (about 22.1GB), using the settings --n-cpu-moe 999, --flash-attn on, -c 32768, -t 16.

By language, JavaScript scored 89.8%, Python 88.2%, C++ 84.6%, Java 76.6%, Go 74.4%, and Rust 53.3%, confirming broad-based improvement. The key takeaway is that agent/scaffold and serving harness significantly determine the performance of local coding models.

benchmarkqwenlocal-llmcoding-agentllama-cppaider-polyglot
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

RedditMica v0.1 4B Released: Open-Source Decision Model for Agent Loops Trained for Under $30TLDR AIQwen Intelligence Launches Three Mobile AI Agents and Open Benchmark SuiteRedditDynamic Quantiser Released, Improving Accuracy Over Standard Quantization in VRAM-Constrained EnvironmentsHacker NewsApple Releases LensVLM-9B for Processing Compressed Text ImagesAnthropic NewsAnthropic Releases Claude Opus 5.5 with 40% Cost Reduction and Record Alignment ScoresTLDR AIQwen Releases 'RecreationWorld' for Training AI Agents to Rebuild Apps