alt Hacker NewsIn my experience, yes. A bit more reliable than gemma for me. I mostly use A3B (35B, mix of experts) though, because it‘s faster, and in the same ballpark intelligence wise as the dense 27B, so it’s the sweetspot for me. I want to try cohere‘s mini code model next, but worried the runtimes aren‘t optimized for that yet.
Replies
regularfry • last Sunday at 12:19 PM
Worth knowing that Unsloth have just put out another Gemma 4 release from Google's upstream updates which should improve reliability. Bugs in the chat template affecting tool calling and other issues, apparently. https://www.reddit.com/r/unsloth/s/MpC6Hzs4Wj
➕ show 1 reply
I found qwen3.6:26b slightly better on my 32G mac mini than the same sized gemma until gemma was updated with better tool support 4 or 5 days ago.
It is like a ping-pong game: the advantage flips back and forth between providers.