VecSzn

don’t be toxic :)

[VecSzn @Vec_Szn]

the repo link is returning 404 right now. GitHub flagged my account today, and that hides the profile and every public repo on it, PolyStrata included. I believe it's a mistake and I've filed an appeal. GitHub doesn't give a timeline for these, so I can't say when it will be back. The code is intact and the project isn't going anywhere. I'll update this comment when the link works again.

Please use hf's mirrir for now: https://t.co/uiH45prHPO

[VecSzn @Vec_Szn]

@QwenDevs We need qwen4 35B MoE

[VecSzn @Vec_Szn]

PolyStrata is built on Strata by Niko1221, and most of the code is that project's: https://t.co/00O3SHUVqe

The four added models are experimental and were measured on two machines. Numbers, method and limits are in the README. MIT license.

[VecSzn @Vec_Szn]

PolyStrata is public: an open-source engine for running large mixture-of-experts models on an ordinary PC.

On a laptop with an 8 GB RTX 4070, Qwen3.6-35B-A3B writes 73-75 tokens/s and reads a 30K-token prompt at 1,390 tokens/s.

https://t.co/ebAIXzXJCI

[VecSzn @Vec_Szn]

It runs Qwen3.8-Flash-Next, GLM 5.3 Flash, Qwen3.6-35B-A3B, Ornith 1.5 and Gemma 4 26B-A4B. One-click install on Windows and Linux, a web app, and an OpenAI- and Anthropic-compatible API on localhost for your coding agent.

[VecSzn @Vec_Szn]

@heroicgg @nicoodoz whats the problem

[VecSzn @Vec_Szn]

@FYBfyb1229 @G2CSGO agreed

[VecSzn @Vec_Szn]

@CounterStrike Waiting for CS2

[VecSzn @Vec_Szn]

@InternetDwarf He really did show some speed

[VecSzn @Vec_Szn]

@G2CSGO Gigahoox