mrfakename releases abliterated DeepSeek-V4.1-Flash tuned to preserve intelligence
realmrfakename · x · 2026-10-18
Known community tuner mrfakename released an abliterated (refusal-removed) version of DeepSeek-V4.1-Flash-1017, emphasizing preservation of the base model's intelligence. The HF page shows directional ablation removes refusal behaviour while keeping the same files, format and FP8 layout as the original, so it loads anywhere the original does. MIT-licensed, with Safetensors and GGUF formats and 8-bit/FP8 quantizations; roughly 763B params.
Related event: mrfakename releases abliterated DeepSeek V4.1 Flash and GGUF quants(5 posts)→
More from Models
- $20 plan showdown: user finds Codex far less token-efficient than Claude — cneuralnetwork · 2026-10-18
- Dave Shapiro ditches ChatGPT's Dot agent over silent failures and opaque model routing — _AustinCalvert_ · 2026-10-18
- OpenAI releases 700+ AI-generated math papers, with claimed progress on Riemann hypothesis — PolarBearby · 2026-10-18
- Steering a single neuron only works for concepts with good pretraining representation — Sauers_ · 2026-10-18
- Every Grok Bot task ends with writing a prompt for local ChatGPT instead — lucasmeijer · 2026-10-18
- DeepSeek V4.1-Flash leaker accused of being a bot: 'not a single MLA layer' — teortaxesTex · 2026-10-18