Looking for suggestions for a local model that would be good at evaluating comedy writing
retained text
I have a specific use case where I have a small web application that prompts the user with a randomly generated premise for a joke (using an llm), and then the user is expected to write a joke with the generated premise as a sort of challenge. An LLM is then used to evaluate their joke as far as how good of a joke they wrote using the supplied premise, provide a score, provide feedback on why it was or was not a good joke, and tips for what they could do to improve. It works well using various large models, and relatively okay using openrouter/free , which has a series of larger free models in a collection and an intelligent router. However I've tested on various smaller models and gotten mixed results. I'm looking for some suggestions from people for models that I may not be aware of that might be good for more general knowledge things like this, rather than coding. I have 64GB of system ram (machine has a discrete GPU / AMD APU so no dedicated VRAM). So I'm probably looking at a 45-70b total parameter model on the higher end, but preferably it's an MOE model with a smaller number of active params for speed. Each use will probably have \~400 tokens input (including system prompt), and \~200-600 tokens output, so I'm okay with heavier models even though I don't have a GPU because it's a small number of total tokens. Some models I've tried, like bonsai-27b and LFM2.5 24b-a2b, have had weird results. Gemma 4 26b-a4b seems to really have a good grasp of it and do the best so far out of any local models I've tried, but I'm wondering if anyone has other suggestions.
document fetched 2026-11-10 00:41 · canonical https://reddit.com/r/LocalLLaMA/comments/1xrv20h/looking_for_suggestions_for_a_local_model_that