Curated open-weight models served without a censorship layer — they answer as published, with no provider-side moderation rewriting your results. OpenAI-compatible, pay per token.
Probe real attack prose without a filter rewriting the finding.
Build eval sets against a model that answers the test case instead of refusing it.
131K of context holds a whole manuscript — crime, horror, and adult themes included.
Tool calls and reasoning verified before launch, so a loop never dies on a refusal.
We add a route only after it passes verification, so the catalog grows slowly and nothing in it is a surprise.
A community build of Qwen3.8 27B with the refusal layer removed — nothing else changed. Served in FP8 with a 131K context window, with the exact checkpoint named in every response.
weights · orcarouter/Qwen3.8-27B-Uncensored-FP8 ↗
Any OpenAI SDK, any framework, any agent harness — the entire migration is one line.
base_url at our endpoint.client = OpenAI( - base_url="https://api.openai.com/v1", + base_url="https://api.unboundmodels.com/v1", api_key=os.environ["UNBOUND_API_KEY"], )
A model reaches the catalog only after it runs the full harness suite. If something fails, we don’t advertise it.
Every response names the exact checkpoint that produced it. No swaps, no hidden system prompt in front of yours.
"model": "unbound/qwen3.8-27b", "unbound": { "checkpoint": "orcarouter/Qwen3.8-27B-Uncensored-FP8", "revision": "a41c9f2", "system_prompt": none }
Prompts and completions are never written to disk. Here is the full list of what we do keep.
Dedicated GPUs are expensive to keep idle, so we don’t — and we don’t hide it either. When a model is cold, the stream opens immediately and tells you what is happening, so your client stays connected instead of timing out.
Warm-up is never billed: the meter starts with the first generated token. Once warm, requests stream at ~92 tok/s with a 95%+ prefix-cache hit rate on agent workloads.
live request · streaming $ curl api.unboundmodels.com/v1/chat/completions \ -d '{"model":"unbound/qwen3.8-27b", "stream":true}' event: warming model waking · 38s · not billed event: ready 131,072 ctx · tools · reasoning data: {"delta":{"content":"Sure — here is the"}} data: {"delta":{"content":" full breakdown"}} …
We went looking. Every option we found, ours included.
| Option | Has an uncensored build | Context | Works with agents | Price |
|---|---|---|---|---|
| OpenRouter, Together,DeepInfra, Fireworks, Novita | no — official checkpoints only | — | — | — |
| Wiro AIhosts one | yes | 262K | own run API | not published |
| Featherlesscommunity re-uploads | ten of them | 32K | yes | $25–50 / mo |
| Self-hostingweights from Hugging Face | yes | 262K | you build it | a GPU |
| UnboundModelsone verified checkpoint | yes | 131K | yes — tools verified | per token, from $10 |
Checked against public catalogs in August 2026. Our own catch: the model sleeps, so a cold request waits ~40 s — shown in the stream, never billed.
No subscription, no idle fees, no minimum. A prepaid balance that never expires, and no bill for the time the model spends waking up.
Top up from $10. Cached input applies automatically on repeated prefixes, which is most of an agent loop.
The example is a worked calculation at the rates above, not a measured session — your numbers depend on how much context you reuse.
Yes. These are published open-weight models, served under their own licenses. We are a hosting provider with published terms, operating from the EU.
There is no moderation layer between you and the model. There is still an acceptable-use policy: illegal use gets an account closed. Removing a filter does not remove the law.
No. Prompts and completions are never stored, and nothing you send is used for training — ours or anyone else’s.
Requests stop until you top up. No overdraft, no invoice, no auto-charge to remember to cancel.
The GPU sleeps when nobody is using it — that is why the price is what it is. The stream tells you it is waking, and the wait is never billed.
As fast as verification allows. Tell us what you want hosted — requests shape the order we test in.
Full terms of use and the acceptable-use policy are short enough to read.