Qwen 3.6 27B vs 35B-A3B: 16GB VRAM Local Test

Qwen 3.6 27B & 35B-A3B Full Review & Tests - Agentic Coding & Local AI Vision Today we dive into Qwen 3.6 mid-size models! In this video, I do a breakdown and extensive testing of the Qwen 3.6 27B (Dense) and 35B-A3B (Sparse MoE) models running entirely locally. We explore their new hybrid architecture combining Gated DeltaNet and Gated Attention, and put them through the wringer with agentic coding, creative writing, vision tasks, and tool calling. These models are not here to play games!! The 27B is even competing with Claude 4.5 Opus on SWE-bench and LiveCodeBench! What you’ll learn in this tutorial: ✅ The architectural breakdown of Qwen 3.6's MoE and Dense models (Gated DeltaNet + Gated Attention). ✅ Generating a complete, responsive dark-theme HTML/CSS/JS website from a single prompt. ✅ Coding a fully functional 3D browser car game from scratch using Three.js. ✅Testing advanced vision capabilities, including complex image-to-code UI recreation and object counting. ✅ Evaluating strict system prompt adherence and creative writing skills. ✅ Setting up and testing Web Search tool calling with Tavily inside Open WebUI. ✅ Dense PDF document reading and extracting specific quotes from heavy technical papers. And so much more!!! Tools & Models Used: Open WebUI: The ultimate frontend for running local models with tool calling capabilities. llama.cpp: For running large GGUF models efficiently locally. Tavily: Search engine API used for the web search tool integration. Unsloth GGUFs: For high-quality, optimized quantized model files. PC Specs: Gpu: Nvidia RTX 5060 Ti 16 GB : https://amzn.to/4rU7xRy Ram: 64gb 4x16gb Kingston Fury : https://amzn.to/473HoaG Model Used : Qwen3.6-27B-UD-Q4_K_XL Qwen3.6-35B-A3B-UD-Q4_K_XL (Paired with the mmproj vision file for both models for hte vision test's) Pro Tip: The Qwen 3.6 27B Dense model is an absolute powerhouse for heavy coding and production tasks it handles complex logic and pixel-perfect UI generation incredibly well! When setting up tool calling, ensure your Tavily API keys are properly configured in the Open WebUI admin panel for seamless web searching. If you found this breakdown helpful, don’t forget to Like, Subscribe, and Hit the Notification Bell for more deep dives into AI-powered coding and local LLMs! ig :   / kintugk   x : https://x.com/gk_kintu Contact: [email protected] Videos Mentioned : GPT 2.0 Image review :    • ChatGPT Images 2.0 -  Full Review and Test   Hermes Agent Tutorial VIDEO :    • Hermes Agent : Full Review and Test   Paddle Ocr VIDEO :    • PaddleOCR-VL-1.5 vs GLM-OCR: Local Test   Gemma 4 VIDEO 31B vs 26BA-A4B:    • Gemma 4 31B vs 26B-A4B: 16GB VRAM Local Test   Timestamps: 0:00 - Intro & Model Overview 0:56 - Benchmarks (SWE-bench & vs Claude 4.5 Opus) 1:41 - Hybrid Architecture Explained (MoE vs Dense) 4:02 - Local Setup (llama.cpp & Open WebUI) 5:43 - Test 1: HTML Website Generation 8:12 - Test 2: 3D Browser Car Game (Three.js) 10:37 - Test 3: Creative Writing (Modern Fiction) 14:09 - Test 4: System Prompt Adherence (Baking AI) 15:23 - Test 5: English to Norwegian Translation 16:46 - Test 6: Dense PDF Document Reading (PaddleOCR paper) 18:37 - Test 7: Vision Tests (People, Glasses & Emoji Counting) 20:48 - Test 8: Web Search Tool Calling (Tavily) 21:50 - Test 9: Image to Code (Admin Dashboard UI) 24:36 - Final Thoughts & Outro #qwen3 #LocalAI #LLM #CodingAI #OpenWebUI #LlamaCPP #MachineLearning #AIWorkflow #AgenticAI