qwen3-4b-opennn

Qwen3-4B for OpenNN

Qwen/Qwen3-4B converted to the flat binary format consumed by OpenNN's examples/qwen3 chat example (grouped-query attention with RoPE + QK-Norm, gated SwiGLU Dense, RMSNorm, tied lm_head).

Both weight files hold bit-for-bit the same model. The upstream checkpoint is BF16, so the FP32 file is those values widened to 4 bytes and the BF16 file is a lossless copy of it at half the size; BF16 is also what OpenNN's CUDA path uses internally. The config and tokenizer files are shared by both.

The example downloads these automatically on first run (build OpenNN with CUDA and run qwen3).

The model needs about 8 GB of VRAM once mirrored to BF16. On machines with less memory use qwen3-0.6b-opennn, the smaller conversion.

Weights are a direct dtype-cast of the original checkpoint; see the base model card for training details and the Apache 2.0 license.

Licence and redistribution notices

Apache License 2.0; Copyright 2024 Alibaba Cloud. The complete original licence is preserved in LICENSE_QWEN3.txt. Keep that file, NOTICE_QWEN3.txt and SOURCE.md with redistributed weights or tokenizer files. These records identify the original Qwen model and Artelnics' conversion and tokenizer adaptations. OpenNN's software licence does not replace the model licence.