Messing around with meme coding with qwen3.5 35b. How does this stack up compared to the paid stuff?
>35bit's trashminimax-m2.5 on ollama is free
>>10835140127B is better.
>>108351401use dense model
I feel as though there is something in this post that has not yet been addressed.
>>108353125what do you reckon that might be, anon
>>108351401>>108353245isn't this miru tights?
>>108353308unfortunately the cosplayer herself is the real mystery...
reverse image search found another one.>>108352881I could never get the bloody moe models to run. which dense?
>>108353375moe models are easier to rununironically just use llama.cpp and ask whatever flavor of corpo LLMs for args with your pc spec loldense i mean 27B models, they will crush 35B-A3B any day for jobs like coding but you need at least 16GB vram for a good experiencemoes are memes
>>108353395yeah i remember trying to get them running with llama.cpp and claude code hooked up to it via some router program. no end of problems. maybe im too retarded, its certainly possible.kept running out of ram on 2x older 24gb cards.Also is all qwen 3.5 just moe? The 27b versions are just quants? Also vision is just wasted data and training, so the code only version is still qwen 2.5 or something?
>>108353528~to 27B it is densefrom 35B-A3B they are all moeand 27B is number of parameters, quants mean they have parameter precision compromised, does not mean lower paramsand just use qwen3.5, not qwen3 coderseriously just run 27b dense Q5 on a single gpu and that will be more than enough for you
>>108353591thanks anon>>108353318picrel
>>108351401>qwen3.5 35b>How does this stack up compared to the paid stuff?Kinda shit I guess? You are using 1–1.5 years old stuff compared to the paid models. You're also missing out on all the most advanced agentic integrations and tools.
