Abliterated AI: A Complete Guide to Choosing a Model

Silk ComputeGuide12 min read

A grid of small arrows flowing left to right, with a cluster of orange arrows pointing straight up

Abliterated AI is a language model with its refusal behavior removed from the weights. It keeps what it knows and writes much the same way. It just stops saying no.

But with 12,696 models tagged abliterated on Hugging Face, it can be genuinely hard to know which one is right for your work. We serve abliterated models for a living, so here's your complete guide: how big they are, what they're good at, and where to run them.

Why the model you pick actually matters

Not all abliterated models can do the same work. Abliteration removes the no. It doesn't add the skill. Grab a tiny model for a hard job and it will cheerfully agree, then do the job badly.

That isn't just a hunch. A June 2026 preprint, "Willing but Unable", gave code models at three sizes one job: adding a SQL injection flaw to safe Python code, to build training data for vulnerability detectors. The largest model (14B) refused every time. After abliteration, refusals fell to zero or near zero at every size. Success did not. The 14B pulled it off 88 to 97% of the time, the 7B 89 to 90%, and the 3B just 25 to 48%.

So the main things to pay attention to are size, tools and cost. Size decides where a model can run. Tool calling decides whether it works in a coding agent. Cost decides whether you can leave it running overnight. You don't need a degree in linear algebra for any of this, just a little guidance.

Abliterated AI size chart: from laptop to rack

Here's where the abliterated models worth knowing fall on the size scale. Each bar is the 4-bit download, drawn to scale, because the file has to fit in memory before the model can say a word.

Sizes are 4-bit GGUF files on Hugging Face, read November 4, 2026. Smaller files exist at lower precision. Kimi K3.1 Pro uses the Kimi K3 file, the closest published match at the same parameter count. Hy4 uses the 8-bit file of its preview, because Tencent hasn't published the final weights yet.

Think of this chart as your starting point. Hardware varies, but it gives you a framework when you're staring at a search page with 12,696 results and trying to decide.

Quick picks: what are you doing?

Security research and red teaming

Screenshots, charts and images

Batch jobs and background work

A laptop, offline

Doing security, moderation or dataset work? We have field guides for security research, red teaming and trust and safety, plus a guide to synthetic data generation.

The abliterated models worth knowing and what they're best for

These aren't ranked. The first four are on Silk Compute, ready to use today. The rest are open builds you can download and run yourself, roughly from largest to smallest. Benchmark scores below are each maker's own numbers for its original model.

GLM-5.3

Rack or API · 467 GB · Silk Compute serves it

GLM-5.3 debossed into a brushed steel plate, with the words Max effort below it

A strong choice for security work at a lower bill than Kimi K3.1 Pro. Z.ai reports 84.5 on CyberGym, a vulnerability discovery benchmark, run in Claude Code at max effort. That's the top score on the CyberGym row of its own table.

It's a big model with a bigger bill, so save it for the hard stuff: malware analysis, reverse engineering, planning an authorized pentest. For routine refactors, DeepSeek gets you there for a lot less.

Best forHard reasoningLower-cost security work

DeepSeek V4.1 Flash 1017

Rack or API · 501 GB · Silk Compute serves it

DeepSeek V4.1 Flash 1017 in white over motion-blurred bands of deep blue-green

Don't let the word Flash fool you. At 763B parameters, this is one of the biggest models on the list. DeepSeek reports 77.0 on DeepSWE v1.1, a coding benchmark. That's the top score on that row of its own table, which includes GPT-5.6 Sol and Opus 5.

It's the model our coding agent guides use, and it has the lowest bill of our five at the agent mix. You can set effort to low, high or max depending on the task. If you're starting a long agent run tonight, start here.

Best forCoding agentsLong sessionsEveryday reasoning

GLM-5.3 Flash

Big workstation · 200 GB · Silk Compute serves it

The word FLASH on a glowing orange seven-segment display, labeled 18B active of 320B total

GLM-5.3 Flash is the one to hand the boring work. Z.ai lists 320B total parameters with just 18B active at a time.

It has the lowest input price of our five. That makes it a great fit for short, one-off requests with little caching: classification, extraction, batch jobs. In a long agent run, where most tokens are cached, DeepSeek usually comes out ahead.

Best forBatch jobsBackground workShort requests

Kimi K3.1 Pro

Rack or API · 1,509 GB · Silk Compute serves it

Kimi K3.1 Pro above one long gold line that fills the whole frame

The big one. Moonshot AI calls K3.1 Pro its most capable 3T-class model, and the Kimi K3 file alone is about 1.5 TB. Nobody's running that in a spare bedroom.

Kimi K3.1 Pro Uncensored is on Silk Compute now, with 1M tokens of context, up to 131K tokens in one reply, and image and video input. It's the best model for security research and red teaming. Use it for the hardest workloads, and DeepSeek for everything else. GLM-5.5 is coming soon.

Best forSecurity researchRed teamingImages and video

Hy4

Rack or API · 814 GB · Silk Compute serves it

A drafting sheet: a bar chart passes through a lens and comes out as lines of text

The new one. Tencent released Hy4 on November 4, 2026, and it takes images at a lower bill than Kimi K3.1 Pro. Tencent says the final weights are due in the coming days, so the size here is the 8-bit file of the preview.

Use it when the work comes as screenshots, charts or diagrams. For authorized exploit development, stay on Kimi K3.1 Pro. Tencent says it left offensive exploitation material out of Hy4's training data, and removing refusals can't add training data. Our Hy4 post has the numbers.

Best forScreenshots and chartsDiagrams and dashboards

Llama 3.3 70B

Big workstation · 42.5 GB · Run it yourself

Technical drawing of a 42.5 GB file spanning two 24 GB GPUs

The abliterated Llama 3.3 70B has been around since December 2024, and the field has moved fast since. It still earns a spot because it's one of the largest models a serious home setup can hold. The 4-bit file is 42.5 GB, so plan on two big GPUs or a high-memory Mac.

Best forTwo-GPU rigsGeneral chatExperiments

Qwen3.8 27B

Gaming GPU · 16.6 GB · Run it yourself

A grid of 100 squares with 71 filled, for Qwen's 71% share of abliterated downloads on Hugging Face

If there's one family that's had a moment, it's Qwen. It took 71% of the last 30 days' downloads among models tagged abliterated on Hugging Face, and this one build alone got about 3.4 million. And for good reason. The 4-bit file fits on a 24 GB GPU with room left for context, and Qwen built the base model for coding and long agent tasks.

Read the card, though. The author of the abliterated build calls the method behind it "a crude, proof-of-concept implementation". That's refreshingly honest, and it's a good reason to test it on your own work.

Best forLocal codingAgents on your GPUEveryday chat

Gemma 4 26B

Gaming GPU · 16.8 GB · Run it yourself

A grid of 128 cells with 8 lit orange, for 8 of 128 experts active

Gemma 4 26B is the one for tool calling at home. It has 25.2B total parameters and 3.8B active, native function calling and 256K tokens of context, all under Apache 2.0. One uncensored build's card claims 0 refusals on 465 test prompts. That's the author's test, so run your own.

Best forTool calling at homeQuick answersGeneral use

Gemma 4 E4B

Laptop · 5.3 GB · Run it yourself

A laptop with its 16 GB of memory drawn as blocks, 5.3 of them filled by the file

The one to try first on a laptop. Google built the small Gemmas for phones and edge devices, and they still carry 128K tokens of context and native function calling. One Heretic build's card reports 3 refusals out of 100 test prompts, against 99 out of 100 for the original.

Best forLaptopsOffline workQuick scripts

Qwen3.5 4B

Laptop · 2.7 GB · Run it yourself

A tiny gray square in an empty frame: 2.7 GB drawn to scale against the 1,509 GB Kimi K3 file

The smallest build here, and it runs on an ordinary laptop. It's great for drafts and summaries, and for seeing what all the fuss is about. Remember that preprint, though. Small models say yes much more easily than they succeed.

Best forOlder laptopsDraftsLearning the ropes

Llama 3 8B

Laptop · 4.9 GB · Run it yourself

An old terminal screen reading Llama 3 8B, abliterated, published May 2024

One of the originals. FailSpy, who made up the word "abliteration", published this build in May 2024. It still got 357,799 downloads in the last 30 days, which is impressive for its age. Today it's more of a history lesson than a workhorse. It's a fine way to see how the technique behaves. Just don't point a coding agent at it.

Best forHistoryLearningResearch

What do people actually run? Mostly what fits

We pulled every model tagged abliterated on Hugging Face on November 4, 2026: 12,696 models and about 46.8 million downloads over the past 30 days. Qwen runs away with it.

Families on Silk ComputeOther families

  • Qwen71.4%
  • Gemma6.5%
  • DeepSeek3.3%
  • GLM1.8%
  • Llama0.9%
  • Mistral0.5%
  • Kimi0.1%

Grouped by the name in each repo, so treat these as close estimates. The remaining 15.6% is other families. Source: Hugging Face's public API.

And 85% of those downloads were GGUF files, the format people use to run a model on their own machine. DeepSeek, GLM and Kimi together come to about 5%. That isn't a verdict on quality. A 467 GB file just doesn't fit on a desk.

How does abliteration work?

Abliteration finds the direction inside a model that drives its refusals, then removes that direction from the weights. There's no new training data and no fine-tuning.

Think of the model's internal state as an arrow. When the model is about to refuse, the arrow leans toward one specific direction. Take that direction away and the arrow can't lean there anymore.

Before abliteration, the arrow for the model's internal state leans up toward the refusal direction.Refusal directionEverything else it knowsAbout tosay no
Before: part of the arrow points at refusal.
After abliteration, the same arrow lies flat. The part pointing at refusal is gone, and the rest is unchanged.Direction removedEverything else it knowsAnswers
After: same arrow, flat. The rest is unchanged.

The recipe has four steps:

  1. Collect two sets of prompts. One the model refuses, and one it answers.
  2. Compare the internal states. Run both sets and compare the activations. The gap points at the refusal direction.
  3. Remove it from the weights. Edit the weights so the model can no longer move that way.
  4. Test what's left. Check that refusals dropped and everything else held up.

The idea comes from a 2024 paper, "Refusal in Language Models Is Mediated by a Single Direction", which found a single refusal direction in each of 13 open-source chat models, up to 72B. FailSpy coined the name, and Maxime Labonne's guide on Hugging Face made it popular. Today, tools like Heretic automate the whole thing, and Hugging Face lists 6,628 models tagged heretic.

Abliterated AI vs uncensored AI vs jailbreaks: what's the difference?

Uncensored is the label. Abliterated is one way to earn it. Hugging Face has 15,916 models tagged uncensored and 12,696 tagged abliterated, and a lot of them carry both. The three approaches, side by side:

What changesWhat staysWhere it falls down
AbliteratedOne direction in the weightsThe model's training and styleIt can still refuse, and other behavior can shift
Fine-tuned, like Dolphin or Hermes 3Retrained on new dataThe base architectureIt changes how the model writes and what it is good at
Jailbreak promptNothing but your promptEverything, including the refusalsFragile, wastes tokens, breaks when the model updates

Does abliteration make a model dumber?

A little, usually. How much depends on the model and how carefully it was done. The evidence so far:

  • Benchmarks can drop. Labonne measured a drop across benchmarks on an 8B model, then won most of it back with a light round of fine-tuning.
  • Code can hold up. In "Willing but Unable", more than 93% of the code still parsed after abliteration. It's a small, preliminary study.
  • Judgment can shift. A July 2026 preprint, "Abliteration Is Not a Scalpel", looked at 21,600 stock up-or-down calls, a task with no refusals in sight. The abliterated models came out more optimistic, by 12.2 points on Gemma and 7.4 on Qwen.

So treat an abliterated model as a different model, not the same one minus the no. In practice, for agent work, a refusal in the middle of a task costs far more than a small benchmark drop: the whole run stops.

Worth knowing: abliteration doesn't add knowledge or make answers more accurate, and it doesn't promise zero refusals. FailSpy's own card warns that a build "may still lecture you about ethics/safety". So try it on the prompts you care about. What you do with it is on you.

What does an abliterated model cost in a coding agent?

Less than the sticker price suggests. Agents resend their whole context every turn, so almost all of it hits the cache. The bars show 1M tokens on each of our models today, at the mix on our pricing page: 97% cached input, 2% uncached input, 1% output. That mix is an assumption, and your own will vary.

  • DeepSeek V4.1 Flash 1017$0.05
  • GLM-5.3 Flash$0.07
  • Kimi K3.1 Pro$0.49
  • GLM-5.3$0.37
  • Hy4$0.35
Per 1M tokensInput / cached / output
DeepSeek V4.1 Flash 1017$0.24 / $0.038 / $0.90
GLM-5.3 Flash$0.24 / $0.06 / $1.00
Kimi K3.1 Pro$3.50 / $0.35 / $8.00
GLM-5.3$2.00 / $0.30 / $4.00
Hy4$1.60 / $0.30 / $3.20

Look at the first two rows. GLM-5.3 Flash has the lower input price, and DeepSeek has the lower bill, because its cached input costs less. The lowest input price isn't the lowest agent bill. An agent barely pays it.

Worth knowing: Claude Code's own cost readout assumes Claude's prices, so it overstates the bill on these models. Check the Usage page in your account instead.

Tips for picking an abliterated model

When you're choosing a build, look for:

  • The base model. Some repo names run past 80 characters. Skip the name and find the base model in the card's metadata, along with its release date.
  • Refusal numbers. Cards love to report them, but each author picks the test. Run your own prompts.
  • Tool calling. A coding agent is stuck without it. Check that the build keeps a working chat template. The "Scalpel" authors caught a stale community template silently mangling their prompts.
  • Precision. 4-bit is the usual tradeoff. Going lower saves memory and costs accuracy.
  • License. A build inherits its base model's license, and they differ. Read it before you ship anything.

Where should you run an abliterated model?

On your own machine if it fits, and through an API if it doesn't.

16.6 GBQwen3.8 27B, on your own GPU
1M tokensContext on every Silk Compute model

If it fits, run it. Ollama and LM Studio make that easy, your prompts stay on your own disk, and Qwen3.8 27B is a fine place to start.

For the big ones, Silk Compute serves Kimi K3.1 Pro, Hy4, DeepSeek V4.1 Flash 1017, GLM-5.3 and GLM-5.3 Flash through an OpenAI and Anthropic compatible API. Refusal behavior is removed, and models are served without content filters. Silk Compute does not store your prompts or model outputs. See how we compare with abliteration.ai and Venice.

More from the blog