Unofficial Qwen3.8-Flash-Next E384 FP4 model trends on Hugging Face
Original title: autotrust/Qwen3.8-Flash-Next-E384-FP4-V2 — new model trending #30 on Hugging Face
The NAS build keeps 384 of 512 experts, cuts weights about 16%, and was merged after GRPO-style LoRA.
autotrust published Qwen3.8-Flash-Next-E384-FP4-V2, an unofficial compact derivative of nvidia/Qwen3.8-Flash-Next-FP4-Local. Neural architecture search keeps 384 of 512 routed experts in every decoder MoE layer and the MTP layer, reducing weights from 102.14 GB to 85.66 GB. A full merged checkpoint after GRPO-style LoRA is included, not an adapter. Single-run scores are mixed versus published base references, and the authors say CVSS scoring remains unreliable.
Unconfirmed: Single-run scores use published base references rather than a rerun, and the authors say CVSS scoring remains unreliable.
- Weights are 85.66 GB versus 102.14 GB for the 512-expert original.
- Active experts stay at 10 plus shared experts; top-10 routing is unchanged.
- Serving was validated with vLLM 0.31 on one B200 using a sitecustomize patch.
- GPQA-Diamond scored 75.76 versus a 77.3 base reference; AIME 2026 scored 95.0 versus 93.3.
- Exact CVSS3 and CVSS4 agreement was about 4.5% and 4.2% on completed answers.
- Organizations
- autotrust
From the original article
3 min read at huggingface.coimage-text-to-text · 42,718 downloads · 2370 likes · license other
By autotrust · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.