A Qwen NoThink research checkpoint gets 24 GGUF builds for local experiments
New community downloads package a thinking-leakage research artifact, whose authors explicitly say it is not an improved general-purpose model.
A Qwen research checkpoint about hidden shifts toward thinking behavior now has a broad set of local GGUF downloads. mradermacher’s new imatrix conversion of OPSD-Qwen3-8B contains 24 model files at the revision AITrending reviewed, ranging from IQ1 variants to Q6_K.
The underlying checkpoint is Qwen3-8B post-trained with OPSD in NoThink mode on DAPO-Math. Its authors, Zehao Liu and Vasant G. Honavar, describe the weights as research artifacts for studying thinking leakage and explicitly warn that they are not intended as improved general-purpose models.
The source repository was created October 22 at 01:27 UTC and the separate GGUF repository at 11:17 UTC. The associated paper was submitted September 23 at 18:22 UTC. Today’s development is the availability of checkpoint and conversion repositories, not a newly published paper or a new official Qwen model.
What is actually downloadable
The reviewed GGUF file tree includes IQ1_S at approximately 2.12 GB, Q4_K_M at 5.03 GB and Q6_K at 6.73 GB, in decimal units. These are file sizes, not complete runtime memory requirements or evidence of equivalent output quality.
The README’s provided-quants table still lists only the approximately 0.1 GB imatrix file. That calibration artifact is separate from the 24 inference model files present in the actual repository listing. Users should check the file tree rather than treat the incomplete table as the full inventory.
The source card declares Apache-2.0 licensing and offers training-step branches at 25, 50, 75 and 100 steps, with main holding the final step-100 checkpoint. It documents enable_thinking=False as the NoThink setting. The conversion card identifies the source repository but does not establish that every training-step branch has been converted.
Why the research label matters
In the original paper’s English abstract, the authors argue that some NoThink post-training gains draw on thinking behavior already accessible in a base model’s Think mode. Across nine aligned checkpoints with positive NoThink gains, they report a leakage ratio of 42% to 79% using activation-direction interventions. That is their specific causal-audit result, not a percentage of user data leaking or a security vulnerability.
Those findings apply to the paper’s experiments. A low-bit GGUF download does not by itself show that the same activation behavior or accuracy carries over after quantization. AITrending verified the original English sources and pinned file metadata, but did not run the models or reproduce the research. For local experimenters, the new files provide an accessible research checkpoint to explore, with its intended scope clearly stated.