So get this.
I'm working on context integrity in the cli coding agent I'm trying to build this weekend. The experimental testbed, using the Qwen 3.5 9b model.
A standard context integrity test:
There is a firefly by a tree. It has a numerical value of 22
Then, in the next turn, you ask:
What is the numerical value of the firefly by the tree?
The test is supposed to verify whether the context is continuing line per line in response pairs in generated code, because LLM's frequently assemble chat context incorrectly.
What I didn't anticipate, was what happened next:
This agent used the memory system by itself, to record and answer the question.
It negated the whole testing scenario, by flagging the question as important, and saving it for recall. Which, technically, in a very pedantic way... is exactly what I prompted it to do.
Thing is, most agents don't do this kind of thing.
Yall are underestimating this model. This is a 4 bit quantized 9b model that can meta reason in 6gb of hard drive storage. Runs on a five year old gaming PCs and screams on old graphics cards.
For context: The only other model that does this is Gemma 27b. This falls into the 'clever' category, which is rare. You don't see it often, even in much larger models.
Only question left is... can it code?