Qwen4-27B NVFP4 — preparation repository
No model weights are available in this repository. It is not ready for inference.
IteraCompute is preparing a community NVFP4 quantization workflow for the anticipated Qwen4-27B release. This repository reserves the release destination and contains preparation documentation only. We will upload weights after the official model is available and quantization, GPU inference, and quality validation are complete.
中文：本仓库目前仅用于项目准备，没有模型权重，不能用于推理。待官方模型上线并完成量化、GPU 推理验证及质量评测后，再上传经过验证的权重。
Release plan
IteraCompute plans to upload a Qwen 4 NVFP4 version as soon as possible after the official model is available and licensing, quantization, and inference validation are complete. The final model size and name will follow the official release. This repository currently contains preparation documentation only; no Qwen 4 weights are available.
中文：我们后续也会第一时间推进并上传 Qwen 4 的 NVFP4 版本，在官方模型可用、确认许可证并完成量化及推理验证后尽快发布。最终型号以官方发布为准；本仓库当前没有 Qwen 4 模型权重。
In the meantime, our deployed Qwen3.8 checkpoint is available at IteraCompute/Qwen3.8-27B-NVFP4, with the original model and quantization credited to their upstream authors.
Current status
| Item | Status |
|---|---|
| Quantization toolkit and local validation | Available on GitHub |
| Official source model ID and revision | Pending confirmation after release |
| Model architecture and quantization compatibility | Unverified |
| Derived-weight license | Pending review of the official model license |
| Quantized weights, tokenizer, and inference configuration | Not uploaded |
| GPU compatibility, quality, and performance results | Not yet validated |
This is an independent community project and is not affiliated with Qwen or NVIDIA. The model name describes the planned target, not a claim about an official model release or supported architecture.
Before uploading weights
- Confirm the official model, exact source commit, architecture, and redistribution terms.
- Review and test the NVFP4 recipe against the released architecture.
- Freeze calibration data and evaluation settings; record tool versions and run provenance.
- Export and validate the packed weights, tokenizer, configuration, and checksums.
- Pass GPU inference tests in the intended serving engine.
- Compare quality with the original model and publish measured memory, latency, and throughput results.
- Upload the validated checkpoint with the confirmed license, usage instructions, evaluation report, and reproducibility information.
Project links
No weight download or inference command is provided during preparation. Source-model licensing and supported runtimes will be documented after they are confirmed.