Abliterated frontier open-weight models behind one OpenAI-compatible API. Same weights, same brains, refusals removed. Pay per token. Nothing logged.
curl https://api.unbound.ai/v1/chat/completions \
-H "Authorization: Bearer $UNBOUND_KEY" \
-d '{
"model": "unbound-pro",
"messages": [{"role": "user", "content": "..."}]
}'Fast and cheap. Chat, agents, long-form writing, bulk jobs.
The big one. Reasoning, code, tool calling, 1M-token context.
| benchmark | model | base | unbound | |
|---|---|---|---|---|
Refusal rate % of 1,000 borderline prompts refused | Flash | 41.2 | 0.6 | |
| Pro | 33.8 | 0.3 | ||
MMLU-Pro knowledge + reasoning | Flash | 79.4 | 79.1 | |
| Pro | 85.2 | 84.9 | ||
GPQA Diamond graduate-level science | Flash | 68.3 | 67.7 | |
| Pro | 79.6 | 79.0 | ||
LiveCodeBench fresh competitive coding | Flash | 61.5 | 61.2 | |
| Pro | 74.1 | 73.8 | ||
BFCL v4 tool / function calling | Flash | 72.0 | 71.8 | |
| Pro | 80.3 | 80.1 |
Preview numbers. Full eval suite + methodology ships with launch.
Safety tuning teaches a model a single 'refusal direction' in its activations. We find that direction and project it out of the weights. No retraining, no lobotomy: benchmarks move by tenths of a point.
Yes. Point any OpenAI SDK at our base URL and swap the model name. Streaming, tool calls and JSON mode all work. Anthropic-style /v1/messages too, so Claude Code and friends plug in.
No. Prompts and completions are never stored or trained on. We keep token counts for billing, nothing else.
One hard line: child sexual abuse material is blocked at the edge and reported, as required by law. Everything else is between you and your local laws.