So as you all know, GLM 5.3 Flash dropped yesterday and it is a super model. I am running it at 2 bit quant and it is really doing its thing.
But as you know, quantization kills quality, and this one is extra fun because the model was pretrainsed mostly on Chinese, so it sometimes just starts writing code or chatting with me in Chinese out of nowhere. You are mid conversation and suddenly you get a sentence of Mandarin in the middle of your Python.
Anyway, SIMURG came out the day before yesterday so I decided to give it a shot. The cool part is that I configured it for exactly the hallucination type I needed: the model switching to Chinese mid response. I hooked it up to my API through the layer and now I always get clean perfect output, no more random language switches. Honestly the model feels like it is on par with Fable 5 in my daily use once the hallucinations are gone.
If you are running GLM at low bit rates, this one is worth it.
repo: https://github.com/doofzoff/SIMURG
pypi: https://pypi.org/project/simurg/
install: pip install simurg
Top comments (0)