distilqwen

SkillAI & models

DistilQwen2.5 — Alibaba's industrial practices for training distilled open lightweight language models. Knowledge distillation from Qwen2.5 72B into smaller 0.5B-7B models.

Available today. Use it from your connected AI after setup.

Connect ahel once, and every AI you use reads what you have installed.

Then ask your AI: use the distilqwen skill

What this skill tells your AI

The instructions your AI receives, as published by mkurman/zorai in skills/scientific-skills/distilqwen/SKILL.md and read by ahel’s review.

Overview

DistilQwen2.5 (Alibaba, arXiv:2504.15027) provides industrial practices for training distilled open lightweight LLMs. The approach distills Qwen2.5-72B-Instruct into smaller models (0.5B, 1.5B, 3B, 7B) with strong performance retention.

Key Techniques

  • Logit-level distillation: transfer output distribution from teacher to student
  • Representation alignment: align hidden states between teacher and student layers
  • Data curriculum: progressive difficulty in training data selection
  • Multi-stage training: pre-training distillation → instruction tuning → preference alignment

Usage

The distilled models are available on HuggingFace as distilqwen/distilqwen2.5-*-instruct and can be used directly:

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("distilqwen/distilqwen2.5-1.5b-instruct")
tokenizer = AutoTokenizer.from_pretrained("distilqwen/distilqwen2.5-1.5b-instruct")

References

Signals

GitHub stars
324
Forks
26
Last commit
Sep 2026
Advanced
Catalog kind
skill
Gateway key
distilqwen
Source
github.com/mkurman/zorai