Abliterated DeepSeek and GLM behind one OpenAI-compatible endpoint

What it does This is the official public repository for a hosted API: six uncensored or “abliterated” DeepSeek and GLM text models served through an OpenAI-compatible chat-completions endpoint. You keep your existing OpenAI SDK client, point it at their base URL, and pick a model ID. There’s also a browser chat if you’d rather not write code. The repo’s actual content is guides, pricing tables, and runnable Python/TypeScript examples — the platform implementation is not included.

The interesting bit “Abliterated” means the model weights were modified to reduce refusal behavior, and the README is refreshingly honest about it: it explicitly says this doesn’t guarantee correct answers or zero refusals. The pricing table is also unusually candid — it warns that catalog inclusion doesn’t mean a model is currently healthy, and points to live performance links instead of pretending everything is fine.

Key highlights

• Six models, all with 1,000,000-token context, streaming, and function tools; output caps range from 65,536 to 130,000 tokens.
• Pricing from $0.30 to $1.40 per million tokens (input and output), with cheaper cached input tiers.
• OpenAI-compatible protocol, so existing SDKs and coding clients (Cline, Continue, OpenCode) work with a config change.
• Docs cover security-research workflows, error/billing semantics (credit reservation before generation), and stream failure handling.
• MIT-licensed docs and examples; hosted in Switzerland; $0.10 trial credits, $3 minimum top-up.

Caveats

• This is glue and documentation around a paid hosted service — the repo contains no server code, so you’re evaluating a vendor, not a tool you can self-host.
• Requests reserve credit before generation, so a small balance may not cover every model; repeat-trial accounts may hit identity verification.
• Rates and availability are subject to change, with the live catalog taking precedence over the README’s snapshot.

Verdict Worth a look if you specifically need refusal-reduced models for security research, reverse engineering, or evaluation work and want OpenAI-SDK compatibility. Skip it if you wanted open weights or a self-hostable stack — the models live on their servers, not yours.