Skip to main content
Back

ReliquaryForge's qwen3-4b-base-dapo-v4 Brings 32B Reasoning Power to a Single GPU

A community-trained Qwen3-4B base tuned with ByteDance's DAPO reinforcement learning recipe hits nearly 900K downloads, showing appetite for small open reasoning models.

  • ReliquaryForge's qwen3-4b-base-dapo-v4 hit 884k downloads and 3.1k likes on Hugging Face.
  • Built on Qwen3-4B-Base: 4B dense, 36 layers, GQA, 32K context, Apache 2.0.
  • Post-trained with DAPO, ByteDance Seed's open RL recipe that hit 50 on AIME 2024 at 32B.
  • DAPO combines decoupled clipping, dynamic sampling, token-level loss, and overlong reward shaping.
  • Ships BF16 safetensors with a chat template; runs in Transformers, vLLM, SGLang, Docker.
  • Requires transformers 4.51.0 or newer to load the Qwen3 architecture.

A community fine-tuned reasoning model built on Qwen3-4B-Base has quietly become one of the most downloaded small models on Hugging Face. qwen3-4b-base-dapo-v4, published by ReliquaryForge, is the fourth iteration of a Qwen3-4B checkpoint further trained with DAPO, the open reinforcement learning recipe ByteDance Seed released earlier this year. It sits at 884k downloads and 3.1k likes, unusual traction for a hobbyist repo with no benchmark table of its own.

The model card itself mostly inherits the upstream Qwen3 documentation, so the interesting story is the recipe and the size class. A 4B dense model pushed through a proper large-scale RL loop is exactly the kind of artifact many teams want for on-device or single-GPU reasoning work, and until recently that combination was mostly locked inside frontier labs.

What's actually in the box

The starting point is Qwen Team's Qwen3-4B-Base, a dense causal LM with 4.0B total parameters (3.6B non-embedding), 36 layers, grouped-query attention with 32 query heads and 8 KV heads, and a 32,768 token context window. Qwen3 as a family was pre-trained on 36 trillion tokens across 119 languages, roughly triple the language coverage of Qwen2.5, with a heavier mix of coding, STEM, reasoning, book, multilingual, and synthetic data.

On top of that base, the v4 checkpoint applies DAPO, short for