That’s up my alley. In my (admittedly not super extensive) experiments, I have found the MoE models with tiny active parameter counts (35b-a3b) to be virtually useless for any real agentic work. Organizing notes in OpenCode scrapes right at the top of their capabilities. Admittedly this is a harsh test with OpenCode’s system prompt being pretty huge.

The dense 27b Qwen models I found to be refreshingly good in comparison. Still not beasts of intelligence, but in my simple tests, but they’d reliably follow instructions and were honestly not that much slower than 35b-a3b! So I’m a fan of larger dense models now for acceptable intelligence.

Further fine-tuning for agentic work certainly sounds like a good thing, but I wonder what they are sacrificing for it. The Readme contains no information about that.

It’s notable that the model comes strongly pre-quantized, making it fit in just over 8GB of RAM. That’s also hugely impressive. I really wonder how it will compare to a less quantized base Qwen.

Also notable: They don’t seem to support reasoning yet, having it disabled by default and treating it as experimental. That ought to tank performance.

The scope appears to be clearly for text-only coding, so I’m fairly certain that the rest of the capabilities is going to suffer. I’m still excited about this type of model because they clearly thought about the needs of local deployment.

Edit: Oh boy, finding out even more: They are using an atypical quantization method called AVQ2 which is not compatible with stock engines, hence they are providing compatibility layers. I’ve seen enough to stop my evaluation here, I think I will wait for engines to catch up. The performance numbers also seem really off - on my Mac, Qwen 3.6 27b is much faster than what they claim (by multiples).