Real-life experiences with Qwen?
retained text
Hello, I read a lot about Qwen here in the last weeks and months. Many people seem to like it, but I'm curious whether anyone here has actually used it for anything else than coding and development. I am looking for an LLM that can help a non-profit org deal with everyday paperwork. It should be able to: * Summarize documents, mails, letters etc. and help draft answers * Use a specified document set for RAG, i.e. answer questions like "How many letters did we get regarding xy, and what did they say?" * Do OCR with both printed and handwritten text * Occasionally do translations to and from french, english and german * Help with creating office documents The last one needs some explanation - our users are mostly more or less computer-illiterate. They know the absolute basics of using LibreOffice, but I would like to help them by either integrating LibreOffice with a local LLM (I already took a look at LibreThinker and LocalWriter) or giving the LLM the tools required to generate OpenDocument files at users request. It should accomplish rather simple tasks such as creating a table with x columns for .... and formulas for creating sums, percentages etc. The last one is a bit tricky as it doesn't depend primarily on the LLM, I know. For the first 4, I already tried Gemma4-26B-A4B and was surprised how good it worked on the testing hardware I have. I currently have an old Server with a Xeon E5-2665v4, 96GB RAM and two Nvidia P106 (6GB each). I don't have any exact numbers yet, but it is fast enough that I would rate it "acceptable". Qwen3.6-35B-A3B seems to be a little bit too much for that hardware, at least it responded more than slowly. I am ready to spend a bit more. Two RTX 3060 12GB would definitely be possible, but that would probably not be enough to run a dense model like Qwen3.8-27B, including the vision projector and a sufficiently large KV cache from VRAM. A single V100 with 32GB RAM or two RTX 5060 Ti with 16GB would probably be in reach, a Radeon R9700 or anything above is too much. Before I do that, I'd like to know if Qwen3.8 is really suitable for that task, if anyone has experiences to share. Thanks!
document fetched 2026-09-01 00:41 · canonical https://reddit.com/r/LocalLLaMA/comments/1w3l8s5/reallife_experiences_with_qwen