SectileLabs/qwen38_27b_nvfp4

Qwen3.8-27B, GInfer NVFP4

This repo holds the GInfer packages of Qwen3.8-27B in NVFP4: single-file packages for the GInfer inference engine (and GChat, which is built on it). They run only on GInfer / GChat. They are not safetensors or GGUF files and other runtimes cannot load them.

GInfer is Sectile Labs' inference engine for fast, high-concurrency, long-context work on NVIDIA GPUs.

Packages

File	Drafter	Tensor-parallel size	Size in bytes	SHA-256
qwen38_27b_nvfp4_df2q4.ginfer	Q4 DFlash2	1 (one GPU)	21,707,145,472	54638e003c17d78f5c049f76d58b24c649786dc9c8b6165ba25b7ab7fe00dd21
qwen38_27b_nvfp4_df2q4_tp2.ginfer	Q4 DFlash2	2 (two GPUs)	21,711,450,112	7cd19361a9694ccd412c2fc95cbe8e1dc491abef942e0c9822d1372fb6862db6
qwen38_27b_nvfp4_df2.ginfer	NVFP4 DFlash2	1 (one GPU)	21,767,307,520	ad4972ce2903d4b16ba6d966ee86a77b8816778832a9d176c26be139de2b77e0
qwen38_27b_nvfp4_df2_tp2.ginfer	NVFP4 DFlash2	2 (two GPUs)	21,771,625,732	637b5556118b5a2ff27267424df6e04c4548e59f0f4b051fb2fecbd291718b04

Format: NVFP4 target weights.

GPUs: NVFP4 needs a Blackwell GPU (RTX 50xx / PRO).

The TP2 packages split the model across two identical GPUs; load them with tensor-parallel size 2. TP4 packages will be added later.

Check each download with sha256sum qwen38_27b_nvfp4_df2q4.ginfer, sha256sum qwen38_27b_nvfp4_df2q4_tp2.ginfer, sha256sum qwen38_27b_nvfp4_df2.ginfer or sha256sum qwen38_27b_nvfp4_df2_tp2.ginfer.

Upstream models and licenses

• Base model: Qwen/Qwen3.8-27B, revision 1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0. License field apache-2.0; the repo LICENSE file is the Apache License 2.0 (full text in LICENSE here).
• Drafter: incoai/Qwen3.8-27B-DFlash2, revision dedf8df68adfb1afeaf7b7480c0a0243108177b4. License field apache-2.0.

Quantized by Sectile Labs from the official BF16 weights of Qwen/Qwen3.8-27B (revision 1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0): the MLP of layers 0-55 to NVFP4 (group size 16), re-rounded with GPTQ on Sectile Labs' own real-text calibration corpus, and the remaining matrices to FP8. The result was converted to a .ginfer package. The DFlash2 drafter (incoai/Qwen3.8-27B-DFlash2) is quantized by Sectile Labs: to Q4 in the Q4 DFlash2 packages and to NVFP4 in the NVFP4 DFlash2 packages.

This package is a conversion of the upstream weights into the GInfer format; GInfer does not change the upstream license. The upstream license text is in LICENSE.