sergqwer / strata-nvfp4

Strata fork: Qwen3.8-Flash-Next 125B MoE in NVFP4 on one RTX 20-50 card (12 GB+) and 64 GB of RAM or more. Setup installs our GPTQ quants (huihui abliterated, OrcaRouter) from Hugging Face. CPU experts on AVX-512 or AVX2 (+AVX-VNNI), all experts in RAM from 92 GB, a low-RAM mode below, vision, W4A8 prompts. Windows / CUDA 13.

Why this job: matched by what the repo talks about: ram, qwen3, flash, moe

Open on GitHub

What it does

Accuracy against Qwen3.8-Flash-Next with every routed expert in Q8_0, speed on an RTX 5090 with 128 GB of RAM. Setup downloads the models it names; each Hugging Face card has its measurements and how to run the files by hand.

Get started

No single install command. Read the steps in the README (opens in new tab)

Maintenance: Active; Last commit 3 days ago
Stars: 235; +12 this week
Licence: MIT; Allows business use
Downloads for: Windows; From the latest release
Latest tag: v0.2.0-beta-nvfp4.3; Released 23 days ago
Built with: C++; 46% of the code

Is it looked after?

Contributors: 2; Last 90 days
Issues: 13% closed; 13 still open
Pull requests: 3 merged; 1 waiting
Releases: 41; 30+ in the last year

Similar repos

Repos similar to this one
Repo	Stars	Last commit	Licence	Downloads for
vllm-project / vllm	96,774	today	Apache-2.0	Mac, Linux
sgl-project / sglang	38,767	today	Apache-2.0	No downloads
kacper-daftcode / vLLM-Moet	573	yesterday	Apache-2.0	No downloads

More: download files, languages, topics and project record

Download files, v0.2.0-beta-nvfp4.3

Windows2 files, 1,649 downloads

All files on GitHub (opens in new tab)

Languages

C++ 46%
Python 28%
Cuda 10%
JavaScript 7%

Topics

blackwell
cuda
llm-inference
moe
nvfp4
qwen
vision
windows
avx2
avx512
gptq

Project record

JobLocal LLMs
StatusListed
Commits5,585
Forks26
Watching7
Stars in 28 days+87
CreatedSep 29, 2026
First seen by usNov 7, 2026
CheckedNov 10, 2026