Qwen3.8-27B-DFlash2

A baseRT drafter sidecar for basecompute/Qwen3.8-27B: the DFlash block drafter (--speculate dflash:), converted from z-lab/Qwen3.8-27B-DFlash2 with basert convert (baseRT 0.3.0). It is not a standalone model β€” it drafts tokens that the target verifies, so the target's output distribution is unchanged (greedy and seeded sampling reproduce the target exactly).

File Size sha256 Notes
Qwen3.8-27B-DFlash2-F16.base 3849185280 3d56eae29f63180d25ebee784483bc04a68123068cbde8526f8c48402cf0390e f16 sidecar, loads on Metal and CUDA
Qwen3.8-27B-DFlash2-Q4.base 1456302080 e6ec0261177cba95a85216a6c62c0388472a40db601555c0fe9b0b58d26606d6 q4 drafter layers (--quantize-drafter), heads f16

Usage

basert pull basecompute/Qwen3.8-27B-DFlash2
basert serve basecompute/Qwen3.8-27B --speculate auto
# or name it: --speculate dflash:basecompute/Qwen3.8-27B-DFlash2

License: the drafter weights are z-lab/Qwen3.8-27B-DFlash2's under its apache-2.0 terms; this repository carries them re-packed for baseRT.

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support

Model tree for basecompute/Qwen3.8-27B-DFlash2

Base model

Qwen/Qwen3.8-27B
Finetuned
(19)
this model