after confirming the big model would not fit i went looking for the ones that would. requirements were simple: uncensored, vision capable, worked in llama.cpp on my amd card, and fit 16 gig of vram. no guesses this time — verified the file sizes on hugging face first.
two real candidates came out of it. the best all-rounder is a gemma 4 26b mixture of experts with four billion active — native vision, tool calling, abliterated so it refuses nothing, and a light quant sits around fourteen gig which fits with kv cache headroom to spare. the closest match to the tweet model is a qwen 3.6 14b abliterated at about nine gig, huge room to breathe.
one caveat that matters on amd hardware. mixture of experts models need every expert resident in memory, so vram tracks the full file size, not the active parameter count. and amd cards need a vulkan build of llama.cpp since there is no cuda path. whisper already runs on the same machine so concurrency is something i have to think about.
the honest framing: uncensored means the refusals are gone, not that the model got smarter. same hallucinations, same limits. but for local experimentation it is exactly the right tool, and one of these is the actual pickup.