• Qwen3.5-27b (BF16) on 2x Pro 6k and Gemma-4-E4B (BF16) on RTX 5090
• Took about 8 minutes total (40k tokens total - but like 10k is opencode prompt)
• One prompt for planning (I answered a few follow ups)
• One shot 1000 lines of code
• Fixed only bug (image preview in chat history) in one go

The chat connects to Gemma-4-E4B-IT running on my workstation via vllm. Qwen had no problems getting all the OpenAI compatibility stuff right.

I may keep using it over 122b-a10b (fp8) for coding, but it's not as good at more creative stuff where the 122b-a10b was an extremely good all-round balance for my setup.

Let's hope they drop a 3.6 of the 122b-a10b.

I like the small Gemma as well. It has strong "small model" vibes, but I can see me using it for "running errands".