DeepSeek's Open Weights Get a Second Life: Shrunk, Uncensored, and Downloaded 30,000 Times a Month

A Hugging Face repo called DeepSeek-V4.1-Flash-1017-Uncensored-GSQ-RCO-GGUF, published by the team behind the local-model app AtomicChat, has pushed past 30,000 downloads a month alongside 228 likes and three active community threads. What's being downloaded is not DeepSeek's model as DeepSeek shipped it. It's a two-step derivative: an extreme low-bit quantization of DeepSeek's 552-billion-parameter, 384-expert V4.1-Flash-1017 model, with its refusal behavior surgically removed on top.

Two separate hacks, one Hugging Face repo

The base model is DeepSeek's own MIT-licensed V4.1-Flash-1017, an October refresh of its Flash line with roughly 552B total parameters, 384 routed experts (six active per token), a shared expert, and a context window north of a million tokens. AtomicChat's repo compresses that down using ISTA-DASLab's GSQ-RCO method, offering rungs as low as 1.75 bits per weight (IQ1_M) that the repo's own documentation puts at anywhere from 121-155GB to 231-266GB depending on which size table on the page you read — small enough, either way, to run on a well-equipped workstation rather than a data-center rack. On top of that compression, the card describes projecting a single refusal direction out of the model's dense residual-writing tensors — attention output, the shared expert's down-projection, and the value projection — at a fixed strength, while leaving the routed experts untouched. The result, AtomicChat reports, is a 0% refusal rate on JailbreakBench at every quantization level; for the IQ2_XXS rung specifically, the card reports GPQA-Diamond essentially unchanged from the stock model, while AIME25 actually moved from 76.7% to 80.0% — a gain, not a wash, and one the card doesn't claim holds across every other quantization level.

Neither half of that pipeline is new on its own. ISTA-DASLab — the lab behind the GSQ-RCO method itself — published a similar compression days earlier of Qwen's 176.9B-parameter Flash-Next coding model, pruning half its routed experts and quantizing the rest down to about 58GB, small enough for 32GB consumer hardware while retaining over 90% of the full-precision SWE-bench score. Refusal ablation, often called "abliteration," has circulated on Hugging Face for small open models for well over a year. What's new is seeing both techniques stacked and applied to a model this large: a 552B-parameter, 384-expert frontier system, repackaged for home hardware within roughly three weeks of its own release.

Why the download curve is bending upward

AtomicChat isn't a solo hobbyist account; it's the publisher of an app built around running open-weight models locally and offline, which turns this repo into a product funnel rather than a one-off release. DeepSeek's decision to ship V4.1-Flash-1017 under an MIT license is what makes all of this legal to redistribute in modified form in the first place — and the combination of permissive licensing with rapidly falling quantization costs is the actual engine behind the week-over-week growth in likes and downloads. Every improvement in extreme-quantization tooling makes a larger slice of frontier-scale open weights viable to repackage this way, and every repackaging adds a data point normalizing the practice for the next, bigger model.

The part the model card doesn't test for

What AtomicChat's card measures is refusal rate and benchmark accuracy — exactly the two things vendors of "uncensored" models typically report, and exactly what a July 2026 preprint argues isn't the right test. That paper ran base and abliterated versions of two mixture-of-experts models, Gemma-4-26B-A4B-it and Qwen3-30B-A3B-Instruct-2507, on thousands of weekly stock up/down calls — a task that never triggers a refusal either way, so any difference has to come from somewhere else. It found abliterated versions were measurably more optimistic than their base counterparts — by 12.2 percentage points for Gemma-4-26B-A4B-it and 7.4 points for Qwen3-30B-A3B-Instruct-2507 — with confidence shifting in opposite directions depending on the model, and no actual improvement in forecasting skill to show for it. In other words, removing a refusal direction didn't just unlock answers the base model declined to give — it changed the model's dispositions in ways a refusal-rate benchmark would never surface.

That gap matters more at AtomicChat's scale than it did in the paper's test models. Nothing in the public record has run that same decision-disposition test on a 552B-parameter, 384-expert model subjected to both extreme quantization and refusal ablation simultaneously — two interventions stacked on top of each other, each with its own potential for side effects, on a model increasingly likely to be deployed as an agent rather than a chatbot. Practitioners downloading this quant to get "DeepSeek's capability minus the filter" are, on the current evidence, getting a model whose judgment may have shifted along with its guardrails — just without a benchmark anywhere that says so.