AtomicChat Strips Refusals From a 763B DeepSeek Build on 12 GB VRAM
AtomicChat released uncensored GGUF builds of DeepSeek-V4.1-Flash-1017, a 763B parameter MoE model. JailbreakBench refusals drop from 85-89% to 0% across all three quantization rungs. MMLU stays within ~1 point of the stock quant, KL divergence around 0.02.
Takeaways
- AtomicChat released uncensored GGUF builds of DeepSeek-V4.1-Flash-1017, a 763B parameter MoE model.
- JailbreakBench refusals drop from 85-89% to 0% across all three quantization rungs.
- MMLU stays within ~1 point of the stock quant, KL divergence around 0.02.
- Runs on a single 12 GB GPU plus 64 GB RAM using Strata , which streams the Engram table from SSD.
- The edit modifies only dense residual writers, leaving ISTA-DASLab's GSQ-quantized experts byte identical.
- Does not yet work in llama.cpp because V4.1's separate Engram file is not supported upstream.
AtomicChat has released three low-bit GGUF builds of DeepSeek-V4.1-Flash-1017 with a weight-level refusal edit. The AtomicChat repository reports zero refusals on JailbreakBench after the edit, limited changes on MMLU-500, and average precision as low as 1.75 bits per weight.
The release combines a 551.9B-parameter mixture-of-experts backbone, which contains 384 routed experts, with a separate Engram lookup table. The model is described as having 552B parameters. Running it on a 12 GB GPU still requires 64 GB of system memory, a fast SSD, at least 341 GB of model storage, and the Strata runtime.
The edit targets residual writers
AtomicChat started from the ISTA-DASLab quantization . Its GSQ-RCO process assigns mixed, non-uniform precision to individual tensors and packages the result as standard GGUF shards. The three published variants average 1.75, 2.0, and 2.25 bits per weight.
AtomicChat left the routed expert tensors byte-identical to ISTA-DASLab’s files. It edited three classes of dense tensor that write into the residual stream, the shared hidden state carried between transformer layers:
- Attention output projections
- Shared-expert down projections
- Engram value projections
The process projects an activation direction associated with refusals out of those weights. AtomicChat applied the edit once at full precision with a reported strength of 2.0, then requantized each modified tensor to its original GGUF type. Preserving the routed experts avoids an expensive end-to-end requantization and keeps every published file the same size as its ISTA-DASLab counterpart.
This technique is commonly called abliteration. It estimates a direction in activation space correlated with refusal responses, then changes selected weight matrices so they contribute less along that direction during generation.
Reporting is indexed from AlphaSignal. Rights remain with the original publisher and cited sources.
Read original report
