Search "uncensored" in LM Studio right now and you'll get 500 results sorted by downloads, and the names at the top look nothing like the uncensored models people were running a year ago. The scene has moved fast: HauhauCS's "Aggressive" line has become the default source for clean, no-refusal Qwen and Gemma builds, DavidAU's multi-stage fine-tunes are pushing small models to benchmark scores that used to require something twice their size, and Mixture-of-Experts architecture has quietly changed what "fits in 8GB" even means.
This guide goes through what's actually sitting at the top of the charts today, what fits an 8GB card cleanly, and what needs a workaround (or a bigger card) to run well.
What "Uncensored" Means With Today's Releases
Most of the current top downloads use one of two approaches, and it's worth knowing the difference because it shows up in the model names:
- Heretic / abliterated models remove the model's internal "refusal direction," the learned pattern that makes it say "I can't help with that." The knowledge and reasoning underneath stay intact; the model just stops declining. DavidAU's "Heretic" releases and the HauhauCS "Aggressive" line both work this way.
- Fine-tuned/merged builds go a step further and retrain or merge the model to also improve capability, not just remove refusals. DavidAU's Fable Fusion series is a good example: it's not just decensored Qwen, it's a multi-stage fine-tune stacked on top of the decensoring.
Most current listings advertise their refusal rate directly, and "0/465 Refusals" has become the de facto standard benchmark you'll see repeated across model cards. It's a genuinely useful signal, though worth remembering it measures refusal rate, not output quality.
The 8GB Reality Check
At Q4-class quantization, a dense model needs roughly 0.55-0.6GB of VRAM per billion parameters, plus overhead for the KV cache. On an 8GB card that puts a comfortable ceiling around 9B parameters for a dense model with a decent context window. Above that, you're either dropping to an aggressive quant that hurts quality, or offloading layers to system RAM, which tanks token speed.
The one major exception is Mixture-of-Experts (MoE) architecture, which is exactly why some of the biggest names on the current download charts, like the 35B-parameter Qwen releases, are realistic on an 8GB card at all. An MoE model only activates a small fraction of its total parameters per token (a "35B-A3B" model activates about 3B per token), so llama.cpp can keep the fast-access parts on your GPU and push the rest to system RAM without the usual speed collapse. This is worth understanding before you write off a "35B" model as too big for your card.
With that context, here's what's actually worth downloading.
Best Pure 8GB Fit: Gemma-4-E4B-Uncensored-HauhauCS-Aggressive
This is the cleanest fit on this list for a straightforward 8GB card. It's a 4B-parameter build of Google's Gemma 4, abliterated by HauhauCS with the same "0 refusals" standard as the rest of their lineup, and it comes with native multimodal support for text, images, and audio. At Q4_K_M it's around 5.3GB, and even the higher-quality Q8_K_P quant only runs about 7.6GB, so it fits with room to spare for context and other applications running alongside it.
It's a 4B model, so don't expect it to out-reason a 9B or 27B pick on complex tasks. But for everyday chat, quick vision/audio questions, and a model that loads fast and leaves your system responsive, it's hard to beat right now.
Good for: Users who want headroom on an 8GB card, multimodal (image/audio) tasks, fast load times.
Best Dense 8GB Pick: Qwen3.5-9B-Uncensored-HauhauCS-Aggressive
If you want the most capability you can get in a straightforward dense model that fits an 8GB card cleanly, this is currently the strongest option on the download charts. It's a 9B parameter Qwen 3.5 build with HauhauCS's aggressive decensoring applied, sitting right at the top of the 8GB comfort zone at Q4 quantization. Expect noticeably better reasoning and instruction-following than older 7B-class uncensored models, without pushing you into offload territory.
A close sibling worth knowing about is Qwen3.5-9B-The-Defiant-Fable-Uncensored-Heretic from DavidAU, which applies his multi-stage fine-tuning approach to the same 9B size class. DavidAU's own comparisons put it just behind his 27B Fable Fusion model in benchmark testing, which is a strong result for something that still fits comfortably on an 8GB card.
Good for: General-purpose daily use, coding help, anyone who wants the strongest dense model that doesn't require any offload tricks.
Best "Punches Above Its Class": Qwen3.6-35B-A3B (MoE variants)
This is where things get interesting for 8GB owners. Both Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive and the community Qwen3.6-35B-A3B-Uncensored-Genesis-Hermes-V7 build are MoE models: 35B total parameters, but only about 3B active per token. Because of that architecture, they're genuinely runnable on an 8GB card using llama.cpp's expert-offload flag (--n-cpu-moe), which keeps the fast-access attention layers on your GPU while pushing less-used expert weights to system RAM.
Real-world reports from 8GB and 12GB card owners running this setup land in a wide but very usable range, commonly 30-50+ tokens per second depending on quant, RAM speed, and how many experts get offloaded, which is a genuinely surprising result for a model this size. It won't match a 35B model running entirely in VRAM on a 24GB card, but it comfortably outperforms what a dense 9B model can do, especially on reasoning and long-context tasks (it supports up to 262K context).
This does take a bit more setup than a simple GGUF download: you'll want a quant like IQ4_XS or Q4_K_M, and you'll need to tune the --n-cpu-moe value for your specific hardware, but it's one of the more rewarding options on this list if you're willing to spend ten minutes tuning it.
Good for: 8GB owners who want dense-model-beating reasoning and are willing to do a little flag-tuning; long-context work.
The Ambitious Pick (12GB+, Not a Clean 8GB Fit): Qwen3.6-27B-Fable-Fusion-711-Uncensored-Heretic
This is currently the single most talked-about uncensored release, and for good reason: DavidAU's multi-stage fine-tune is the first model in its size class to break 700 on the ARC-C benchmark at both 8-bit and 4-bit, a threshold previously associated with closed frontier models, and it outperforms the base Qwen 3.6 27B model on 6 of 7 benchmark categories.
Being straightforward about it: this is a dense 27B model, and dense 27B doesn't fit cleanly in 8GB of VRAM without dropping to an aggressive IQ2-class quant, which meaningfully hurts output quality. If you have a 12GB or larger card, it's worth the download. On 8GB, you're better served by the 9B dense pick above or the 35B-A3B MoE option, both of which give you a genuinely good experience without fighting your hardware. Keep this one bookmarked for your next GPU upgrade.
Good for: 12GB+ cards; skip on 8GB unless you're comfortable with heavy quality tradeoffs at IQ2.
Getting Set Up
All of these are standard GGUF files and load the same way in LM Studio: search the model name, pick a quantization your VRAM supports (LM Studio's compatibility indicator will flag quants that won't fit), and download. For the MoE picks specifically, LM Studio's GPU offload slider handles the expert-offload behavior automatically in recent versions, so you don't need to touch the command line to benefit from it, though power users running raw llama.cpp will want the --n-cpu-moe flag for finer control.
Once a model's running on your PC or Mac, LMSA lets you carry that same uncensored setup in your pocket, connecting from your Android phone over Wi-Fi or remotely through Tailscale, so whichever model you land on isn't locked to your desk.
The Bottom Line
For most people on an 8GB card right now, Qwen3.5-9B-Uncensored-HauhauCS-Aggressive is the strongest straightforward pick, no offload tricks required. If you want more reasoning power and don't mind a bit of setup, the Qwen3.6-35B-A3B MoE builds genuinely deliver more than their file size suggests. And if multimodal support or a lighter footprint matters more to you, Gemma-4-E4B-Uncensored is the cleanest, fastest-loading option on the list. The 27B Fable Fusion model is worth knowing about even if it's not a fit today, since it's a good marker of where this next GPU upgrade could take you.