SAVRN
Search Contact SAVRN

Open-weight model

minsik-repro

by MINSIK CHOI CHOI0126/minsik-repro

This is an unfinished training run, published as it goes. Checkpoints appear here every half epoch of a planned six. Nothing here is a final result, and the numbers below will move.

Parameters
Context
Weights117.7 GB
Licenseother
AccessOpen weights
Monthly Downloads

Model Card

This is an unfinished training run, published as it goes. Checkpoints appear here every half epoch of a planned six. Nothing here is a final result, and the numbers below will move. A DFlash draft head for Qwen/Qwen3.6-27B, trained as a reproduction of DFlash (arXiv:2602.06036) on a model the paper does not cover. Full method, scripts and measurement records: It is not a standalone model. It cannot generate text by itself. It writes a 16-token block in one forward pass, conditioned on the target's hidden states at layers 1/16/31/46/61, and the target then verifies that block in a single pass and commits the leading run that matches. Greedy verification means the committed tokens are exactly…

Excerpt from the card by MINSIK CHOI, licensed other.

Identity and Version

Repository
CHOI0126/minsik-repro
Publisher
MINSIK CHOI
Task
Not stated by the source
Modality
Other
Library
Not stated by the source
Parameters
Not stated by the source
Languages
Not stated by the source
Revision
45d2dfb7d6d4001271035df2d8df8c50e511d5e3
First published
2026-09-14
Last updated
2026-09-18

Files and Weights

38 files, 117.7 GB in total. The weights are 26 files totalling 117.7 GB in pt, safetensors.

Weights26 files · 117.7 GB
Configuration10 files · 11.7 KB
Documentation1 file · 6.2 KB
Repository1 file · 1.6 KB
Every file
FileTypeSizeSHA-256
epoch-1.50-step-144174-full/training_state.ptWeights3.5 GB fd69ceaebab7
epoch-1.50-step-144174-full/training_state_rank0.ptWeights5.2 GB 4610d9aed83a
epoch-1.50-step-144174-full/training_state_rank1.ptWeights5.2 GB b9ce9aea1d11
epoch-1.50-step-144174-full/training_state_rank2.ptWeights5.2 GB b159497be702
epoch-1.50-step-144174-full/training_state_rank3.ptWeights5.2 GB 7d14467288e9
epoch-1.50-step-144174/model.safetensorsWeights3.5 GB ef12e4f0ef33
epoch-2.00-step-192232-full/training_state.ptWeights3.5 GB 18f229952bc0
epoch-2.00-step-192232-full/training_state_rank0.ptWeights5.2 GB 856a3bed8883
epoch-2.00-step-192232-full/training_state_rank1.ptWeights5.2 GB eb9245fc478e
epoch-2.00-step-192232-full/training_state_rank2.ptWeights5.2 GB 1ac48b924935
epoch-2.00-step-192232-full/training_state_rank3.ptWeights5.2 GB eaaa763542e3
epoch-2.00-step-192232/model.safetensorsWeights3.5 GB 77e683e4f548
epoch-3.00-step-288348/model.safetensorsWeights3.5 GB 23e2f56cf221
epoch-3.50-step-336406/model.safetensorsWeights3.5 GB d8e39789b0bb
step-48058-full/training_state.ptWeights3.5 GB 9daa06b22cd1
step-48058-full/training_state_rank0.ptWeights5.2 GB 1fa67cc396f0
step-48058-full/training_state_rank1.ptWeights5.2 GB 9c3c0cdc78f9
step-48058-full/training_state_rank2.ptWeights5.2 GB 7f7bb2943888
step-48058-full/training_state_rank3.ptWeights5.2 GB bef0f9c9bb3b
step-48058/model.safetensorsWeights3.5 GB d5e57932bde9
step-96116-full/training_state.ptWeights3.5 GB 00adbd41d10a
step-96116-full/training_state_rank0.ptWeights5.2 GB 96cae85e07d1
step-96116-full/training_state_rank1.ptWeights5.2 GB c392b637ad97
step-96116-full/training_state_rank2.ptWeights5.2 GB b7b0ce7191d4
step-96116-full/training_state_rank3.ptWeights5.2 GB 92cf09eedb27
step-96116/model.safetensorsWeights3.5 GB 3506d2265ac3
epoch-1.50-step-144174-full/RESTORE.jsonConfiguration1.1 KB
epoch-1.50-step-144174/config.jsonConfiguration1.2 KB
epoch-2.00-step-192232-full/RESTORE.jsonConfiguration1.1 KB
epoch-2.00-step-192232/config.jsonConfiguration1.2 KB
epoch-3.00-step-288348/config.jsonConfiguration1.2 KB
epoch-3.50-step-336406/config.jsonConfiguration1.2 KB
step-48058-full/RESTORE.jsonConfiguration1.1 KB
step-48058/config.jsonConfiguration1.2 KB
step-96116-full/RESTORE.jsonConfiguration1.1 KB
step-96116/config.jsonConfiguration1.2 KB
README.mdDocumentation6.2 KB
.gitattributesRepository1.6 KB

License and Download

License
other
Access
Open weights, no gate
Download size
117.7 GB
Download from MINSIK CHOI

Released by MINSIK CHOI through its official repository on Hugging Face.

Built From

Memory Requirements

PrecisionWeights in memory
As published117.7 GB

Weights only, from the published parameter count; the key-value cache and runtime add to this.

Questions About minsik-repro

What license is minsik-repro released under?

other, as its publisher declares it. Read the license text before commercial use.