In the release page: "The other model(s) in Qwen3.8-series would be released later", we know that Qwen3.8-27B is dropping next, but that statement implies there might be other models beyond the 27B!
this post was submitted on 12 Aug 2026
51 points (98.1% liked)
LocalLLaMA
5041 readers
22 users here now
Welcome to LocalLLaMA! Here we discuss running and developing machine learning models at home. Lets explore cutting edge open source neural network technology together.
Get support from the community! Ask questions, share prompts, discuss benchmarks, get hyped at the latest and greatest model releases! Enjoy talking about our awesome hobby.
As ambassadors of the self-hosting machine learning community, we strive to support each other and share our enthusiasm in a positive constructive way.
Rules:
Rule 1 - No harassment or personal character attacks of community members. I.E no namecalling, no generalizing entire groups of people that make up our community, no baseless personal insults.
Rule 2 - No comparing artificial intelligence/machine learning models to cryptocurrency. I.E no comparing the usefulness of models to that of NFTs, no comparing the resource usage required to train a model is anything close to maintaining a blockchain/ mining for crypto, no implying its just a fad/bubble that will leave people with nothing of value when it burst.
Rule 3 - No comparing artificial intelligence/machine learning to simple text prediction algorithms. I.E statements such as "llms are basically just simple text predictions like what your phone keyboard autocorrect uses, and they're still using the same algorithms since <over 10 years ago>.
Rule 4 - No implying that models are devoid of purpose or potential for enriching peoples lives.
founded 3 years ago
MODERATORS
you are viewing a single comment's thread
view the rest of the comments
view the rest of the comments
Yes. Generally, the larger the model, the more knowledge they have.
However, knowledge is not everything, 27b (even 3.6) is much better than Qwen 3.5 at calling tools and general programming.
27b is also a dense model (27b parameters are active at the same time) while 122b is a MoE model with just 10b active parameters. Usually this makes 27b much smarter at understanding complex situations.
I would recommend trying both and seeing which performs better in your tasks. You can also use something like llama-swap to seamlessly swap between them.
If you need something fast, Qwen3.6-35b is also MoE with 3b active parameters. It is really fast and fairly competent in many fields.
Gemma-4-31b is another dense model that does really well with languages (translations, summaries, ...). There's not really a one size fits all with local models but with swapping them around you can get some great results, especially on a Framework Desktop where multiple of the smaller ones might fit at the same time.
FYI: llama-server has built-in model swapping now so you don't need a second piece of software (at least for the simple cases). I pass
--model-presetsand a path to an INI file with the per-model parameters when I run the server.Thanks, my knowledge on the topic of Local AI is still very barebones. I generally use my FW for coding assistance, troubleshooting Linux issues, automating stuff and doing some heavy text processing (fixing artifacts after OCRing old PDFs and formatting text). I have lots of spare RAM as I'm lucky enough to have the 128GB model I bought before the prices went crazy. I need to try this Qwen3.6 and Gemma as you recommended as it looks like they may fit my needs as well. The Coder I have has been running smoothly for a while and I didn't really have the opportunity to tinker with other options recently. If you have any tips for my use case I'm all ears!
128GB is plenty to try different models and see what works best. Have fun!
In case you didn't see it yet, your best friend for Strix Halo machines is https://strixhalo.wiki/
Lots of info on how to setup llama.cpp (with pre-made toolboxes), configure the BIOS settings for maximum available VRAM and Linux for the best performance.
I would recommend setting um llama-swap together with the Strix Halo Toolboxes with llama.cpp. So you can freely swap between models.