Generated: 2026-05-23 · API: Gemini 2.5 Flash · Modes: Summary
Qwen 3.6 27B vs 35B Local AI Agents: Anki Translation Performance
Clip title: Testing Qwen 3.6 35B vs 27b Locally on a Project Author / channel: Jarods Journey URL: https://www.youtube.com/watch?v=ykqIQ6TFohw
Summary
The video provides a comparative test of two new Qwen 3.6 large language models (LLMs), the 27B and 35B variants, using them as local AI coding agents. The main objective is to evaluate their performance in adding a new feature—English translation of Japanese sentences—to an existing Anki flashcard add-on. The speaker aims to determine if these open-source models can serve as effective substitutes for commercial LLMs like GPT-4 or Claude Opus in a practical development workflow.
The testing methodology involved setting up separate Git branches for each model (35B and 27B) within an “OpenCode” environment, which acts as an interface for the AI agent. Both Qwen models were launched locally using VLLM for optimized inference. The speaker then used an add-feature skill in OpenCode to guide each LLM through the process of understanding the project, creating a feature plan, and implementing the translation functionality. This new feature was designed to automatically generate English translations for Japanese sentences and integrate them into the Anki flashcards.
During the evaluation, both Qwen models successfully implemented the desired translation feature, but with distinct characteristics. The Qwen 3.6-35B model was significantly faster in generating responses and code but exhibited a less “intuitive” interaction, prompting fewer clarifying questions and at times feeling more like a “Codex model” that just implements. It initially encountered an HTTP 400 error during API calls, which the speaker resolved by adjusting the API implementation to use standard Python libraries, thereby bypassing VLLM endpoint compatibility issues. The Qwen 3.6-27B model, while approximately three times slower in its processing, proved to be “smarter” and more “intuitive.” It asked more probing and relevant questions regarding the configuration structure and potential overwrites, demonstrating a deeper understanding of the existing add-on’s architecture and potential impacts. It also encountered an initial error, but similar to the 35B, it provided sufficient debugging steps for resolution.
In conclusion, both Qwen 3.6 models showcased impressive capabilities in assisting with coding tasks. The speaker chose the Qwen 3.6-27B model for his personal workflow due to its superior intelligence and proactive clarifying questions, prioritizing accuracy and alignment with his project’s design over raw speed. He noted that while the 35B’s speed could facilitate faster iteration for quick fixes, the 27B’s thoughtful approach led to a more robust implementation from the start. The video highlights the rapid advancement of local LLMs, suggesting they are quickly catching up to the performance of closed-source models from just a few months prior, and emphasizes that quantized versions are available for users with less powerful hardware.
Video Description & Links
Description
Links referenced in the video: Qwen 3.6 collection - https://huggingface.co/collections/Qwen/qwen36
Laptop that I use: Lenovo Thinkpad P1 Gen 7 - https://bit.ly/4haIt46
Hardware for my PC: Graphics Card - https://amzn.to/3pcREux CPU - https://amzn.to/43O66Ir Cooler - https://amzn.to/3p98TwX RAM - https://amzn.to/3NBAsIq SSD Storage - https://amzn.to/42NgMFR Power Supply (PSU) - https://amzn.to/430bIhy PC Case - https://amzn.to/447499T Mother Board - https://amzn.to/3CziMXI
Alternative prebuilds to my PC: Corsair Vengeance i7400 - https://amzn.to/3p64r22 MSI MPG Velox - https://amzn.to/42MnJHl
Cheapest and PC recommended: Cyberpower 3060 - https://amzn.to/3XjtZoP
Come join The Learning Journey! Discord - https://discord.gg/Mym3MxcvWg Github - https://github.com/JarodMica TikTok - https://www.tiktok.com/@jarodsjourney
If you found anything helpful, please consider supporting me and the content I am trying to produce! https://www.buymeacoffee.com/jarodsjourney
URLs
- https://huggingface.co/collections/Qwen/qwen36
- https://bit.ly/4haIt46
- https://amzn.to/3pcREux
- https://amzn.to/43O66Ir
- https://amzn.to/3p98TwX
- https://amzn.to/3NBAsIq
- https://amzn.to/42NgMFR
- https://amzn.to/430bIhy
- https://amzn.to/447499T
- https://amzn.to/3CziMXI
- https://amzn.to/3p64r22
- https://amzn.to/42MnJHl
- https://amzn.to/3XjtZoP
- https://discord.gg/Mym3MxcvWg
- https://github.com/JarodMica
- https://www.tiktok.com/@jarodsjourney
- https://www.buymeacoffee.com/jarodsjourney
Related Concepts
- Local AI Agents — Wikipedia
- Large Language Models — Wikipedia
- Qwen LLMs — Wikipedia
- Translation Performance — Wikipedia
- Anki Flashcard Add-on — Wikipedia
- Code Generation — Wikipedia
- LLM Benchmarking — Wikipedia
- VLLM Inference — Wikipedia
- Open Source LLMs — Wikipedia
- Model Parameter Scaling — Wikipedia
- Anki Flashcards — Wikipedia
- Python Development — Wikipedia
- Git Workflow — Wikipedia
- API Integration — Wikipedia
- Quantized Models — Wikipedia
- Software Engineering — Wikipedia
- Debugging Strategies — Wikipedia
- Prompt Engineering — Wikipedia