Q4 and Q5 versions please?

#35
by nazeshinjite - opened

Can you make Q4 and Q5 versions for 128GB systems? It would be amazing to have your quality for a popular hardware size.

https://huggingface.co/ISTA-DASLab/Qwen3.8-Flash-Next-GSQ-RCO-GGUF#results
Do you think it will turn out even better than the original?

believe it or not but real work is usually not a benchmark and usually takes more perceived impact than a benchmark score does.

On codding tasks it struggles a lot more than Q5-Q6-Q8, this is why you don't see any SWE-* benchmarks in their comparison table. Frankly speaking thier bench table is very cherry picked. They literally picked tests which don't degrade with model quality loose.

IST Austria Distributed Algorithms and Systems Lab org

Sorry for the late reply, everyone! I'm starting my PhD and we're also working on improving our methods so things have been pretty busy here and I haven't been able to respond to comments for a while.

@nazeshinjite
Based on the community feedback we've received and our internal testing, we really think that going higher than this would bring diminishing returns. If your hardware allows you to run a larger model, I'd recommend trying multiple agents in parallel or using a larger context size for a better experience.

@ladno64
I should clarify this. We would never claim that a quantized model is better than the base model. What you're seeing is just the mean of three benchmark results with high variance. The difference is essentially noise and IQ3_S and the base model perform similarly from a statistical standpoint.

@auf1r2
We only run a few small benchmarks that can be completed relatively quickly. We'd love to include SWE benchmarks as well, but running them would take several days per model and we'd much rather spend that time improving our methods and working on more models given the limited computational capacity that we currently have. Even the benchmarks we provide are relatively uncommon among the major labs releasing quantized models. Usually, you can only find a few KLD numbers which take just minutes to calculate.

We're always transparent about the limitations of our models, and so far, we haven't seen any major degradation in performance compared to larger models. For the best experience we recommend using xhigh reasoning with the default generation parameters.

Based on community testing, the IQ3_S model matches the base model on long-horizon agentic coding tasks with the same token usage:
https://x.com/bnjmn_marie/status/2107660419880063161

Thank you for the reply. But I think not only I, but also many users of the Strata project would be grateful to you for a model with a higher bit depth, at least one at 4.5 bits. Please!
images (2)

yea, i know posted benchmarks are saying iq3 is performing good. but i just cant get myself to use anything less than q4 for anything. we all have been trained for how long now to know how fast models drop after q4 with q4 being the sweat spot. so for now i have to stick with the orca standard iq4 quant.

Sign up or log in to comment