Qwen3.6 35B-A3B<\/a> only activates a fraction of its parameters per token – typically 3B out of 35B total. The model has 256 experts, picks 8 per token, and only those 8 fire.<\/p>\nThe result: you get the reasoning quality of a 35B model with the VRAM footprint closer to a 3B model. On 8GB VRAM, this is the unlock that makes everything else possible.<\/p>\n
The 9 Models I Tested<\/h2>\n
I tested models across four categories: small fast workers, medium daily drivers, large reasoning models, and one failure. Here’s the lineup:<\/p>\n
Small Fast Workers<\/h3>\n
Gemma 4 E4B<\/strong> (Q4_K_M quantization)<\/p>\n\n- Google’s small MoE model – 4B active parameters<\/li>\n
- Ran at 64k context, CUDA backend<\/li>\n
- This was the speed champion of the entire benchmark<\/li>\n<\/ul>\n
Qwen3-4B Thinking<\/strong> (Q4_1 quantization)<\/p>\n\n- Alibaba’s small thinking model<\/li>\n
- Designed for reasoning tasks despite its size<\/li>\n
- Good for simple planning tasks<\/li>\n<\/ul>\n
Medium Daily Drivers<\/h3>\n
Qwen2.5-Coder 7B<\/strong> (Q4_K_M quantization, 128k context)<\/p>\n\n- The dedicated coding model from Alibaba<\/li>\n
- 7B dense parameters – fits comfortably in VRAM<\/li>\n
- This is what I reach for most coding tasks now<\/li>\n<\/ul>\n
Qwen3.5 9B<\/strong> (Q4_K_M quantization)<\/p>\n\n- General-purpose model with vision capabilities<\/li>\n
- Tested at 125k context – it worked but the context pressure was real<\/li>\n<\/ul>\n
Large Reasoning Models<\/h3>\n
Gemma 4 26B-A4B<\/strong> (Q4_K_M and UD-IQ4_NL quantizations)<\/p>\n\n- Google’s larger MoE – 26B total, 4B active per token<\/li>\n
- Tested at both 4k and 64k context with different MoE layer offload counts<\/li>\n
- Solid reasoning, slower generation<\/li>\n<\/ul>\n
Qwen3-Coder 30B-A3B<\/strong> (IQ4_XS quantization)<\/p>\n\n- Alibaba’s large coding MoE – 30B total, 3B active<\/li>\n
- Tested extensively with MoE layer offload from 20 to 41 layers<\/li>\n
- Interesting but not the daily driver I hoped for<\/li>\n<\/ul>\n
Qwen3.6 35B-A3B MTP<\/strong> (UD-IQ4_NL quantization)<\/p>\n\n- The most advanced Qwen model with multi-token prediction (MTP)<\/li>\n
- Tested at 32k and 128k context<\/li>\n
- MTP with draft-n=1 was the best 32k performer<\/li>\n<\/ul>\n
Ornith-1.0-35B-MTP-APEX-I-Mini<\/strong> (UD-IQ4_NL quantization)<\/p>\n\n- Qwen3.5 MoE variant – 35B total, 3B active, 256 experts (8 active per token)<\/li>\n
- The model I ultimately settled on<\/li>\n
- Stable at 128k context with turbo2 KV cache compression<\/li>\n<\/ul>\n
The Failure<\/h3>\n
GLM-4.6V-Flash<\/strong> (Q4_K_M + mmproj)<\/p>\n