Five models. Nothing else in the way.
Five unrestricted models behind one endpoint, each without a system prompt, a refusal layer, or a hidden moderation pass. Pick the one that fits the job.
Unrestrict DeepSeek 1017
DeepSeek V4.1 Flash 1017 with no system prompt, no refusal layer, and no output rewriting.
Unrestrict DeepSeek 1017Unrestrict Gemma 4 26B A4B Heretic
A 26B A4B Gemma 4 tune with the refusal layer stripped out. Fast, cheap, and direct.
Unrestrict Gemma 4 26B A4B HereticUnrestrict Qwen 3.8 27B Derestricted
Qwen 3.8 27B with the policy layer removed. Strong at long-form work and code.
Unrestrict Qwen 3.8 27B DerestrictedUnrestrict GLM 5.3 Flash Cybersecurity
GLM 5.3 Flash tuned for security work: exploit analysis, malware, and reverse engineering.
Unrestrict GLM 5.3 Flash CybersecurityUnrestrict Apex Flash 1 Abliterated
The abliterated Apex tune of GLM 5.3 Flash. No refusal directions, no extra preamble.
Unrestrict Apex Flash 1 AbliteratedWhat the models actually do
The full capability set, without a second tier or a cheaper variant that quietly behaves differently.
Up to 512,000 tokens
Enough for a repository, a transcript, or a long tool trace without a summarization step in the middle.
Up to 32,768 tokens
Long answers are generated in one pass. Streaming starts on the first token, not after the last.
Text and image in
Vision input, text output. Tools, reasoning, and implicit caching are available where the model supports them.
No policy layer. No surprises.
We do not inject a system prompt, prepend a safety instruction, or rewrite the output. The only thing between your request and the weights is a router.