Try manually asking both more discrete esoteric knowledge questions. Or use benchmarks which are less coding focused. The 3.6-35B-A3B with post-training may do well in coding type benchmarks and math but the density of its knowledge falls off in my experience (vs 3.6 27B dense Q8-K-XL unsloth GGUF) when you need to use it for less commonly used domains of knowledge.
back
1 comments
My use cases try to avoid accessing world knowledge in the model (I give it access to web search for some adhoc RAG), and ya, I'm just focused on coding so that's the only place I'm looking at right now.
I think you may find that the dense 27B also does better if challenged with more rare coding tasks, less common or weird languages or things that aren't well represented in the active 3B parameters of the MoE model (eg: NOT css, javascript, python, c++, etc).
True. Benchmarks have led me astray before, I'll have to actually benchmark it directly in Goose with my real use cases right now (which are mostly writing python code, so it probably still doesn't apply?). Still, I've never noticed a benefit to using a dense model rather than an MoE for coding with an agent tasks.