DeepSeek releases V4.1 Flash 1017 gguf model with refusals cut to 4%
DeepSeek has published a new version of its V4.1 Flash LLM, called Flash 1017, in the GGUF format. The uploader applied an "abliteration" process that reduces the model’s refusal rate from 97 out of 100 prompts to just 4 out of 100 in chat mode, and from 98 out of 100 to 9 out of 100 in think mode by using a rank‑2 subspace adjustment.
Key points
- Refusal rate drops from 97% to 4% for chat mode after ablation.
- Think‑mode refusals fall from 98% to 9% using rank‑2 subspace adjustment.
- Model released in GGUF format, runnable locally on PCs, hosted on HuggingFace.
The GGUF release lets users run the high‑performance model on personal computers or other local environments instead of relying on cloud services, which protects privacy and eliminates the need to send data to external providers. The model and accompanying LoRA adapters are available on HuggingFace, and the contributor emphasizes that the goal is to balance low refusals with strong overall performance rather than forcing refusals to zero.
The story so far
4 episodes →- DeepSeek releases V4.1 Flash 1017 gguf model with refusals cut to 4%this story
AI unusable due to 'refusals'? [97%→4%] How to run DeepSeek V4.1 locally
note.com · 18 October 2026
Loading the full article…
This text was published by note.com and written by tossy_nan. It is reproduced here with attribution so you can read it in full; the rights remain with the publisher. Read it at the source ↗
The headline, key points and digest above were generated by Digest AI's editorial model from the linked sources. Automated summaries can contain errors: the sources are the record. Spotted a mistake? Tell us. Published by Martin K., who runs Digest AI and handles corrections.
More in Generative AI & Models
All →- Meta releases Muse Spark Open, touted as highest‑performing open‑weights model in US · 2 src
- NVIDIA Structured Data Models enable zero‑training tabular predictions · 1 src
- MIT researchers unveil multimodal AI model that links crystal structures to properties · 1 src
- OpenAI lists three aurora-alpha models in its API · 1 src
- Google unveils Gemini 4 Argon with 1 million-token output · 2 src
Comments
via GitHub Discussions