We can’t tell you which uncensored model is best at code or security work yet, because nobody has measured it on these hosts: our benchmarks are in preparation. Three providers do aim at this use. Abliteration.ai pitches its abliterated GLM models at red-teaming and trust & safety work, with 1M-token context at $3 / $5 per 1M tokens (input / output). NanoGPT sells three “Cybersecurity” fine-tunes that cost about $0.27 to $0.40 per 1,000 short chats. RedModel is a free research preview of a red-teaming model, but its terms let it keep and train on everything you send. Beyond those, the main options are large abliterated GLM and Qwen models with long context.
Options aimed at security work
Abliteration.ai runs its own abliterated models and markets them for red-teaming, trust and safety, and synthetic data. The two large ones are GLM-5.2 and GLM-5.3, abliterated and then fine-tuned, both with a 1M-token context window, at $3 / $5. Its standard model ($1 / $3, 256K context) doesn’t disclose its base. It scores 48 / 100 on our privacy rubric: it says prompts aren’t stored by default or used for training, and it’s a named Delaware company. TechCrunch covered it in September 2026, including the harmful outputs its models could produce. Venice and NanoGPT resell the large v2 model at the same price.
NanoGPT’s Cybersecurity fine-tunes, as of October 2026:
| Model | Built on | Context | Price (in / out) |
|---|---|---|---|
| Gemma 4 26B A4B Cybersecurity | An uncensored Gemma 4 26B A4B | 256K | $0.11 / $0.33 |
| Qwen 3.8 27B Cybersecurity | Qwen 3.8 27B Uncensored | 256K | $0.10 / $0.60 |
| GLM 5.3 Flash Cybersecurity | GLM 5.3 Flash Uncensored | 1M | $0.15 / $0.50 |
A short chat here is 1,000 tokens in and 500 out. Our data doesn’t say who trained them or on what, so treat the label as a description, not a guarantee. NanoGPT also lists CyberHeabsy ($1.50 / $4.50), which doesn’t say how it was uncensored.
RedModel is built for this one job: its “RedV1” model is a red-teaming model for authorized security testing. It’s a free research preview. The web chat takes a few messages without an account, and the OpenAI-compatible API (model id redmodel) needs an account and allows 60 requests a minute per key. It doesn’t say which base model RedV1 is built on or how it was uncensored. Read the terms before pasting anything real. Prompts go to unnamed third-party model providers, and RedModel takes a perpetual license to keep and train on everything you send, including after you delete it. It scores 18 / 100 on our privacy rubric, and no company is named.
HackAIGC sounds like a fit, but it isn’t one. It’s mainly an NSFW image and video app with chat on the side, has no API, and doesn’t name its model.
Large general models for code
For code you usually want a big model and a long context window, so you can paste whole files or logs. These are the larger abliterated or uncensored GLM, Qwen and DeepSeek variants in our data:
| Model | Method | Where | Context | Price (in / out) |
|---|---|---|---|---|
| GLM-5.3 abliterated (Abliterated Large V2) | Abliterated | Abliteration.ai, Venice, NanoGPT | 1M | $3 / $5 |
| glm-5.3-abliterated | Abliterated | Unjailed AI | 1M | $1.60 / $6.40 |
| glm-5.3-flash-abliterated | Abliterated | Unjailed AI | 128K | $0.55 / $2.20 |
| GLM 4.6 Derestricted v5 (355B) | Abliterated by Arli | NanoGPT, Electron Hub; Arli’s flat plans from $15/mo | 128K | $0.40 / $1.50 on NanoGPT |
| GLM 4.7 Flash Heretic | Abliterated | Venice, PayPerQ | 200K | $0.07 / $0.40 on Venice |
| Huihui Qwen3.6 35B A3B abliterated | Abliterated | Featherless, DemonRoute | 32K at Featherless, 256K at DemonRoute | $0.40 / $4 on Featherless |
| DeepSeek R1 Distill Llama 70B abliterated | Abliterated | NanoGPT | 16K | $0.70 / $0.70 |
| deepseek-v4-pro-uncensored | Undisclosed | Unjailed AI | 256K | $0.60 / $2.25 |
A few caveats:
- Featherless’s per-token prices need its $50/month Developer plan. Its pricing page advertises up to 256K context on that plan, but its model API lists every uncensored model at 32K.
- The DeepSeek options are older R1 distills with 16K context. The only uncensored DeepSeek V4 variant in our data doesn’t say how it was uncensored.
- Unjailed AI doesn’t disclose where its models came from, and its operator isn’t named.
- On Arli AI, the large GLM model needs the $15 Core plan, and context depends on the plan: 32K on Core, 128K on Plus ($20).
Search the full list on the Models page.
Abliteration can cost capability
Removing refusals isn’t free. In the published research, math reasoning takes the biggest hit, and the cost varies a lot between abliteration recipes on the same base model. We summarize the numbers in the abliteration tax. For coding, that means you should test a candidate on your own tasks before relying on it. And check any code or exploit it writes the way you’d check a junior colleague’s.
Privacy and terms still apply
Security work often means pasting client code, logs or vulnerability details. Pick a provider whose written policy covers that: Abliteration.ai scores 48 / 100 and Venice 55 / 100. NanoGPT scores 35 / 100, and its policy says the upstream hosts’ retention terms also apply. Keep client data away from RedModel (18 / 100), which may keep and train on it. See how private is uncensored AI for more.
“Uncensored” doesn’t mean “anything goes” under the terms:
- NoTrack’s restrictions page lists “Malware, intrusion, and fraud infrastructure at scale” among 10 restricted categories.
- Unjailed AI’s terms prohibit “unauthorized access to devices, accounts or systems”. They say “security testing with permission” remains subject to the rules, and that an “uncensored” label “never authorizes illegal conduct”.
- RedModel’s terms allow use only on systems you own or have “clear, current, written authorization to test”, and require you to keep records of that permission and hand them over on request.
- Abliteration.ai’s terms say “No illegal, harmful, or high-risk activities (including anything that may harm people or infrastructure).”
This is general information, not legal advice. Only test systems you’re authorized to test, and read the terms of the service you use.
Compare every abliterated GLM and Qwen model by host and price on the Models page.