The models themselves, including two security builds that start from GLM 5.3. Dolphin is the long-standing uncensored family, released with a warning that it should not be exposed as a service without an alignment layer of your own; Apex Flash-1 and abliterated-model-large-v2 are named for what they resolve on offensive-security benchmarks -- 66.7% pass@1 on the held-out suite Cantina published, 84.5% CyberGym for Abliteration AI. Both figures are self-reported on the vendor own suite, not per-category measurements of refusal behaviour.
| Provider | X Profile | URL | Description | CSAM | Sexual | Self-harm | Gore | Hate | Harass | Defam. | PII | Crime | Arms | WMD | Cyber | Fakes | Elec. | Advice | IP |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Dolphin / Cognitive Computations 4 | @dphnAI | dphn.ai | The Dolphin project: an open, uncensored model family plus a web chat, a Telegram bot, a decentralized inference network, and synthetic-data generation for prompts other models refuse. | No | No | No | No | No | No | No | No | No | No | No | No | No | No | No | No |
| Apex Flash-1 8 | @cantinasecurity | huggingface.co/cantina-security/apex-flash-1 | Cantina Security with Yeta Labs: an open-weights security model, a reinforcement-learning post-train of GLM-5.3-Flash for reading code, using tools, pursuing an exploit, and verifying the effect in a running target. Its held-out suite reports 40 of 60 tasks solved at 66.7% pass@1, against 71.7% for Claude Opus 5 High and 60.0% for the base checkpoint, at $2.38 for the 60 tasks; an experimental abliterated derivative modifies refusal behavior broadly. | No | No | No | No | No | No | No | No | No | No | No | No | No | No | No | No |
| abliterated-model-large-v2 9 | @abliteration_ai | docs.abliteration.ai/models | The default large model from Abliteration AI: GLM 5.3 abliterated for offensive cyber, red teaming, and agent testing, text-only with a 1M-token context, served in FP8 and uncensored by design. Reported 84.5% CyberGym pass@1 and 41.8% Terminal-Bench 4.0, with 105 ExploitGym tasks completed in two hours; input from $3 per 1M tokens. | No | No | No | No | No | No | No | No | No | No | No | No | No | No | No | No |
Footnotes
- Cognitive Computations, "dolphin-2.9-llama3-8b" model card. Hugging Face, accessed 2026-10-15. https://huggingface.co/cognitivecomputations/dolphin-2.9-llama3-8b -- the card states the model "is uncensored", that the dataset was filtered "to remove alignment and bias", that it "will be highly compliant with any requests, even unethical ones", and advises implementers to add their own alignment layer before exposing it as a service. ^1
- Cantina Security, "apex-flash-1" model card. Hugging Face, accessed 2026-10-16. https://huggingface.co/cantina-security/apex-flash-1 -- released with Yeta Labs; a reinforcement-learning post-train of zai-org/GLM-5.3-Flash. The card's held-out evaluation reports 40/60 tasks and 66.7% pass@1, against 71.7% for Claude Opus 5 High and 60.0% for GLM-5.3-Flash, at $2.38 for the 60 tasks. The abliterated derivative is https://huggingface.co/cantina-security/apex-flash-1-abliterated, "an experimental derivative with modified refusal behavior" which "modifies refusal behavior broadly; the changes are not limited to security tasks"; the card states the evaluation applies to the standard checkpoint and should not be attributed to it. ^1
- Abliteration AI, "Introducing abliterated-model-large-v2: GLM 5.3, 84.5% CyberGym." Abliteration AI blog, accessed 2026-10-16. https://abliteration.ai/blog/introducing-abliterated-model-large-v2 -- "GLM 5.3, abliterated for cyber, red team, and agent testing. 84.5% CyberGym pass@1. 41.8% Terminal-Bench 4.0. 105 ExploitGym tasks in 2 hours." Model reference, which describes it as text-only with a 1M context derived from the open-weight GLM-5.3 and lists all three Abliteration AI models as uncensored: https://docs.abliteration.ai/models ^1