[Lwerewolf]

Pretty sure this might be a duplicate. Regardless, tried the 1bit bonsai 27b gguf three different ways - their llama.cpp fork (prism, was it) on 2 machines (1255u/16g, m5 max/128g) and the web (on the 1255u). With llama.cpp it works. on the web it emits the same thing over and over. Prompt was "do you see anything wrong with this code: <C code, compound literal passed as a pointer to a function in an if statement>", result was literally "Yes, let'ssomestruct_struct_tsomestruct<a bit of similar garbage>structstructstruct<forever>.

Locally, it's definitely not the full 3.6 27b, but for ~6G with the context, it's quite impressive, and I like what this implies for larger models. Speeds on the 1255u are abysmal, though (~1TPS or less) - granted, CPU-only.

[kees99]

Qwen is quite slow in general.

I'm getting ~0.5 tps from [qwen], and ~10 tps from [gemma] - roughly the same size, same quant, same hardware (8-core CPU) same software (llama.cpp).

[qwen] Qwen3.6-27B-Q4_K_M.gguf

[gemma] gemma-4-26B-A4B-it-qat-UD-Q4_K_XL.gguf

[christkv]

Quite different models though Qwen 3.6 model has 27B active parameters. gemma-4 has 26B parameters with 4B active at any given time. You should compare with Qwen 3.6 35B A3B