Oriel — AI & technology worth noticing, every day.

autotrust/Qwen3.8-Flash-Next-E384-FP4-V2

AI model · https://huggingface.co/autotrust/Qwen3.8-Flash-Next-E384-FP4-V2

First seen by Oriel on .

AI summary

This page documents an unofficial, fully merged Qwen3.8-Flash-Next-E384-FP4-V2 checkpoint trained with three GRPO-style LoRA updates, its vLLM serving setup, benchmark results, and CVSS limitations.

  • The repository contains all 21 weight shards, tokenizer, configs, and MTP files, with the final LoRA adapter already merged into the model.
  • Serving was validated with vLLM 0.31 on one B200 and requires the included vllm_patch/sitecustomize.py for the checkpoint’s FP4/CSF/PLE layout.
  • Against base references, the model scored 95.0 on AIME 2026, 98.78 on HumanEval, 87.37 on MMLU, and 77.88 on tool calling.
  • On 120 AIME samples, accuracy increased from 92.5 to 95.0 while average completion tokens fell 2.05%, largely through fewer long-tail truncations.
  • Training-prompt CVSS completion improved substantially, but exact agreement remained about 4.5% for CVSS3 and 4.2% for CVSS4, so scoring is unreliable.

#qwen #vllm #reinforcement learning #model quantization

Generated by gpt-5.6-sol. Check the original for details.

Metric trends

Last 30 days · Up to 120 recorded samples per source and metric, with reposts kept separate · First and latest retained

Likes

Total likes reported by the source.

  • Hugging Face Models: 2,415, latest observation ; +545 over observed period since

Downloads (30d)

Rolling 30-day downloads. A change in this value is not new downloads during the observed period.

  • Hugging Face Models: 52,593, latest observation ; +21,880 over observed period since

Aliases

SourceExternal IDURL
Hugging Face Modelsautotrust/Qwen3.8-Flash-Next-E384-FP4-V2https://huggingface.co/autotrust/Qwen3.8-Flash-Next-E384-FP4-V2 Read discussion