Great post with some useful insights!

According to

reddit.com/r/LocalLLaMA/comments/1we8tl1/3827b_has_ruined_353635bs_for_me_its_just

, Qwen3.8-27B (dense) is, unsurprisingly and as can be seen in the previously mentioned artificialanalysis evaluation, a superior model VS Qwen3.6-35B-A3B for most tasks. So, if you are (really) ok with waiting many times longer for a much better answer (for simpler tasks, Qwen3.6-35B-A3B can be good enough and is much faster), you should consider huggingface.co/unsloth/Qwen3.8-27B-GGUF (Q4_K_XL). Due to also being smaller (27B vs 35B), Qwen3 27B has the advantage of giving you many more context tokens[1] on a 32 GB RAM-only device.

The slow speed of running a 27B dense quant on a DDR5-only device is ultimately going to annoy you, so, as previously mentioned, you should really think about getting a device with an at least 8 GB VRAM GPU, so that you can offload parts of the LLM into the much faster VRAM and run a ~27B dense, 4-bit quant, (much) faster VS a DDR5-only device.

[1]
Context: 16,384[KV cache context tokens per GB][2]*(32[GB]-6[GB for OS]-[quant filesize]):

• ~59,000 context tokens when using Qwen3.6-35B-A3B-UD-Q4_K_XL.gguf = 16,384*(32-6-22.4) (or ~26,000 context tokens when giving the OS 8 GB, instead of 6 GB, this is how much every GB matters)
• ~138,000 context tokens when using Qwen3.8-27B-UD-Q4_K_XL.gguf = 16,384*(32-6-17.6)

In other words, Qwen 27B more than doubles your available context tokens to approx. 76,800 (16,384*(22.4-17.6)).

[2] reddit.com/r/Qwen_AI/comments/1vo8pjz/qwen3827b_kv_cache_works_out_to_64_kibtoken_so/