Qwen3.6 27B vs Gemma 4 31B: Memory Recall Battle with a Single Winner

👉 Run these AI benchmarks with me (it's free): https://www.protorikis.com 🎬 👉 Next episode - where everything comes together:    • Qwen3.6 Solves a Brutal Reverse Engineerin...   In this video, I test whether Gemma 4 31B dense model is able to recall context better than its MoE counterpart Gemma 4 26B A4B - the one I tested in the previous episode. And whether the dense Qwen3.6 27B can get an ideal benchmark run! I built a custom benchmark that challenges local LLMs to reproduce exact lines from a massive JavaScript source file (8,705 lines / 336KB / ~108K tokens). The test is brutal: verbatim recall of function bodies positioned anywhere in the file. The results are exciting, but also with important failure points. You'll learn: How to design a positional recall benchmark for long-context models- How Sliding Window Attention in Gemma 4 works Is Gemma 4 31B is better at recalling context than 26B A4B and why How Qwen3.6 27B architecture saves memory for context compared to Gemma 4 31B Whether Gemma 4 and Qwen3.6 dense models are better at context than their MoE counterparts Models tested: Gemma 4 31B Q4 K M unsloth Qwen3.6 27B Q4 K M lmstudio-community Gemma 4 26B A4B UD Q4 K XL unsloth Referenced videos: 👉    • Qwen3.6 vs Gemma 4: Which Actually Remembe...   - (Previous) the Gemma 4 26B A4B vs Qwen3.5 and Qwen3.6 35B A3B Memory Recall Benchmark Hardware for reference: MacBook Pro M3 Max 36GB 🤝 Business inquiries: [email protected] ⏱️ Chapters 00:00 - Intro 00:45 - Memory Recall Benchmark Challenges 01:35 - Gemma 4 31B Architectural Advantage 03:10 - Gemma 4 31B Benchmark Try 1 03:47 - Gemma 4 26B A4B vs 31B Context Memory Usage 04:47 - Qwen3.6 27B Test & Gate DeltaNet 05:31 - The Positional Query 07:31 - Gemma 4 31B Benchmark Try 2 09:27 - Conclusions