Qwen3.6 35B A3B

“Qwen3.6 35B and other finetunes have similar behaviour to Minimax M2.7. It's like a slightly worse M2.7.”

“It made it. But all the graphics were squares and circles.”

“Q4-ish quant of Qwen3.6 35b a3b would be my starting point with that machine. MLX is no longer king by default, however, try oMLX and llama.cpp (via LM Studio etc) and see what happens.”

“Yet my Q4_K_S Qwen3.6 35B A3B cant even write the simplest C++ HTTP server. I dont understand what im doing wrong”

“the other 3 models you mentioned aren't going to be capable of coding anything even close to the game OP was showing”

“The model itself isn't that capable, the quantization reduces accuracy/increases hallucinations, and KV cache quantization further reduces precision.”

“I had a M4 Pro with 48GB, and it wasn't quite enough for e.g. Qwen3.6 35b a3b or 27b with large context.”

“when I hook this model to the background agent workload, it was far worse than the well trained Qwen 3.6 35B-A3B”

“the 35B model reviews it, runs tests, and optimizes it for better performance, readability, and a smaller footprint”

“Each is a step up”

“Qwen 3.6 35b actually gets the job done but boy does it come off dry”

“standard Qwen3.6 35b”

“you had to tell them several times just to fix one problem”

“Their 36b-a4b beats qwen3.6-35b-a3b on all but one of those benchmarks”