Qwen 3.6 27B Neo Code Q4 KM I matrix is badass
Posted by Available_Hornet3538@reddit | LocalLLaMA | View on Reddit | 14 comments
So i am using this model in tax accounting. Have a shitty Ryzen 9 7940HS (8C/16T), 60 GB RAM, Radeon 780M iGPU, 1 TB Kingston NVMe, Win 11 Pro.
It is slow but long story short accurate. It creates excel files from PDF tax forms and other excel sheets. It is interesting as Claude does same thing faster but haven't seen any quality difference. It is just speed. I think this is the next phase of local models running industry specific tasks vs just all coding. Don't need coding just need smarts. Seems like coding models have smarts though.
dead_dads@reddit
Yo! New to local LLMs/ai stuff in general. I have an old 3090 and 128gb of DDR4 RAM. Was going to sell my old machine for parts but occurred to me this week I could turn it into an ai machine to dip my toes into locally run stuff.
My interest rn is to work on some vibe coding projects. Would like to assess and test models that fit fully into the VRAM of the 3090 but also curious about utilizing my ram (DDR4) to see what larger models can bring into the equation.
What models would be worth by time for testing? I’ve been working with Claude to ID some stuff of interest but as this field moves so fast I thought asking people who are actively engaged in this stuff would be better.
Available_Hornet3538@reddit (OP)
To update i have templates the model updates and also a md file the model references --sort of like claude skill. i was thinking about vectorizing the templates and skill file but haven't tested yet. Qwen 3.6 27b is surprisingly smart especially fine tuned on coding tasks. Sort of second nature accounting. I tie everything out, foot and cross foot. So far it works. LM Studio and TUI you can find on github called Late.
Pleasant-Shallot-707@reddit
What’s the benefit of this version over the official model?
Downtown-Pear-6509@reddit
hi. i have similar without 8845hs please teach me good sir!
FatheredPuma81@reddit
People and their obsession with David's shitty "finetunes".
LetsGoBrandon4256@reddit
Model name be like
FatheredPuma81@reddit
At least he didn't "expand" the model to 40B like his "Opus 4.6" finetunes. Or finetune it on a tiny Opus 4.5 dataset and say call it Opus 4.6.
Pablo_the_brave@reddit
And write something about Q2 better than Q8 😃
maxpayne07@reddit
What agent are you using? I have using opencode not only for code but also for office productivity. I maybe try hermes this week. Btw, ryzen 7940hs is not shitty, its a beautifull IGPU 😁😁
Pablo_the_brave@reddit
Are yo use Vulkan with that 780M? Can you share your context and speed?
Enough_Big4191@reddit
yeah we’re seeing similar, speed is nice but correctness on structured tasks matters more. the tricky part shows up when it has to map entities across files, one wrong match and everything downstream looks “fine” but is off. are u validating outputs against a source of truth or mostly eyeballing results right now?
Evanisnotmyname@reddit
Yeah this shit scares me.
I’ve been working on 20 pages of documents and while even working on grounding through hybrid RAG it’s still crazy how quick things lose accuracy.
I’d highly recommend looking into specific models for accounting and finance…which still are inaccurate.
OneSlash137@reddit
Lmfao. Trusting a lobotomized model with tax papers. What could possibly go wrong?
ATK_DEC_SUS_REL@reddit
You’re a cunt.