← sources

Qwen3.8-Flash-Next on Strata is such a beast

source: reddit (reddit:StableDiffusion)published 2026-10-02 16:06ingested 2026-10-03 02:38

retained text

For all your local-LLM needs (prompting, images descriptions...) **Qwen3.8-Flash-Next** is a great model. It's a 180B params model ("*Number of Parameters: 125B with 6B activated, plus 51B n-gram embedding and 4B MTP*" as per their own description) that can run on a beefy home computer, but not absurdly so, by using Strata.

* The source model : [https://huggingface.co/Qwen/Qwen3.8-Flash-Next](https://huggingface.co/Qwen/Qwen3.8-Flash-Next)
* An uncensored quantization : [https://huggingface.co/orcarouter/Qwen3.8-Flash-Next-Uncensored-GGUF](https://huggingface.co/orcarouter/Qwen3.8-Flash-Next-Uncensored-GGUF)
* Strata project : [https://github.com/Niko1221/Strata/](https://github.com/Niko1221/Strata/)

What Strata does that other LLM local hosting method don't :   


https://preview.redd.it/pzgus7skq2th1.png?width=960&format=png&auto=webp&s=ed3b1ce41d4e2899de236b9dfe6871a22fc9403b

With this I ran *Qwen3.8-Flash-Next-Uncensored-IQ3\_XXS* on my Windows 10, **64GB RAM**, 5070Ti **16GB VRAM**, at around **40-50 t/s** with **256K** max context (but running with 32-64K context is more than enough to run it as an image assistant).  
  
If you run their `docs/AI_SETUP.md` through an agent, be sure to point it toward the `docs/ORCA.md` doc too, so it can adapt it. Also ask it to add the `mmproj` used by the other models so it has VL capabilities. 

It handles loading and unloading the model if you want to launch ComfyUI between two prompting sessions. It takes around 2mn to load.

Note : Without Strata you can still run this model directly on llama.cpp with `--cpu-moe` option if you wish, but expect around halved performances, and a noisy CPU fan.

document fetched 2026-10-03 00:32 · canonical https://reddit.com/r/StableDiffusion/comments/1wvxq9n/qwen38flashnext_on_strata_is_such_a_beast