Access powerful abliterated AI models through a single, developer-friendly inference API. Built for agents, automation, security research, and businesses that need greater control.
Live as a limited private beta. Applications are reviewed and access is granted in batches.
client.chat.completions.create()
partner inferencebetaadditional providersroadmapdedicated GPUsroadmapOne key and one endpoint in front of external inference partners. Multi-provider routing, automatic failover and dedicated capacity are on the roadmap — not available at beta launch.
Three abliterated models with refusal behaviour removed, served behind the same API: fast reasoning for agent loops and volume, efficient general-purpose inference, and a larger model for advanced reasoning. Every endpoint is one model parameter away from the next.
model: "model-zero"Fast, low-cost reasoning tuned for agent loops, tool calls and high-volume pipelines.
model: "model-one-flash"Efficient general-purpose inference for everyday production traffic.
model: "model-one"Our larger model for advanced reasoning and long, complex tasks.
Rates are launch prices in USD per 1M tokens, paid from prepaid credits. Cached input is prompt tokens served from cache. Context windows and speed are published per model only after they're measured on production infrastructure — no numbers until then.
HummingRelay speaks the OpenAI Chat Completions format. If your stack already talks to OpenAI, migration is a base URL, a key and a model name.
Set base_url to the HummingRelay endpoint. Your existing OpenAI client library does the rest.
A single API key covers every model in the catalog. Keys are issued to approved beta accounts.
Swap model between Model Zero, Model One Flash and Model One without touching anything else.
Same response object, same usage fields — so your parsing, logging and cost tracking keep working.
import os
from openai import OpenAI
client = OpenAI(
base_url="https://api.hmbrd.dev/v1",
api_key=os.environ["HUMMINGRELAY_API_KEY"], # hr_live_...
)
response = client.chat.completions.create(
model="model-zero",
messages=[
{"role": "system", "content": "You are a security analyst on an authorized engagement."},
{"role": "user", "content": "Review this nginx config for misconfigurations: ..."},
],
)
print(response.choices[0].message.content){
"id": "chatcmpl-…",
"object": "chat.completion",
"model": "model-zero",
"choices": [
{
"index": 0,
"message": { "role": "assistant", "content": "1. server_tokens is on …" },
"finish_reason": "stop"
}
],
"usage": { "prompt_tokens": …, "completion_tokens": …, "total_tokens": … }
}Parameters, streaming, billing and every error code are in the API documentation →
HummingRelay is the layer between your code and the models: auth, model selection, billing and — over time — routing and privacy controls. Here is exactly what the beta has today and what comes after.
Specialised models with refusal behaviour removed at the weights level — for workloads where blanket refusals break the job.
One base URL, one key, one request format. Point your existing OpenAI SDK at HummingRelay and change the model string.
Top up credits and pay only for the tokens you use, at published per-model rates — one balance instead of one bill per upstream.
Direct onboarding for companies: custom volume, procurement questions and dedicated deployment discussions.
Requests are routed across partner inference providers by availability, performance and cost, so you are not tied to one upstream.
Transparent fallback to another provider when an upstream degrades — no retry logic in your code.
Optional middleware that detects, redacts or tokenises personal data before a request leaves HummingRelay for an external provider.
Self-hosted inference on dedicated hardware for customers who need isolation or guaranteed throughput.
Abliterated models remove blanket refusals so legitimate work stops getting blocked. Accounts are still bound by our Acceptable Use Policy: no illegal activity, no attacks on systems you are not authorised to test, no abuse of third parties. Access is reviewed, and keys are revoked for violations.
Build autonomous workflows that need consistent model behaviour — without a tool-calling loop stalling on an unnecessary refusal halfway through a run.
Support authorised penetration testing, defensive research and security automation, where mainstream models routinely refuse to discuss the very thing you are defending against.
Integrate specialised models into internal systems behind one vendor relationship, with configurable access and privacy controls as they roll out.
Evaluate model behaviour, study alignment and refusal mechanics, and compare abliterated models side by side through one interface.
No subscription and no monthly fee. Pick a one-off top-up pack — the bigger the pack, the bigger the bonus — and every request is deducted at the published per-model rates. During the beta, top-ups are applied by our team; there is no self-serve checkout.
Once per account — a small first top-up to try the API.
One-off top-up, no subscription.
One-off top-up, no subscription.
One-off top-up, no subscription.
Any other amount from $1 works too: +15% bonus from $20, +20% from $100, +30% from $500; no bonus below $20.
The bonus part is bonus credit: spent first, valid on Model Zero and Model One Flash, not on Model One. Top-ups are final: no refunds, no withdrawals. Credits do not expire monthly.
Message our team on Telegram and it lands on your balance. Credits stay there until you use them.
Every request is metered in input, cached input and output tokens and deducted at that model's rate. Bonus credit is spent first, except on paid-only models.
You pay HummingRelay only — never a separate bill per inference provider.
HummingRelay launches as a limited-access service. Every application is reviewed, and access is provided gradually as capacity comes online.
Abliteration is a technique that identifies the internal direction a model uses to represent refusal and removes it from the weights. The result keeps the base model's knowledge and capabilities but stops issuing reflexive, blanket refusals.
It is not a jailbreak or a prompt trick — the change is in the model itself, so behaviour is consistent.
Mainstream APIs serve aligned models that refuse whole categories of legitimate work — security research, red-teaming, frank analysis. HummingRelay serves abliterated variants behind one OpenAI-compatible endpoint, with a single key and a single prepaid balance across models and providers.
Yes — it follows the OpenAI Chat Completions format, so the official OpenAI SDKs work by changing the base URL, API key and model name. Streaming, tool calls and the supported parameters are covered in the API documentation.
We're starting with three abliterated models: Model Zero, fast reasoning for agent loops and high-volume work; Model One Flash, efficient general-purpose inference; and Model One, a larger model for advanced reasoning. Per-token rates are listed in the model catalog; context windows and speed are published once measured on production infrastructure.
There is no subscription. You top up prepaid credits with a one-off pack — Just Try, Prime, Ultima or Scale — or any other amount, and each request is metered in tokens and deducted from your balance at the published per-model rates for input, cached input and output. Larger top-ups carry a larger bonus; see Pricing. During the beta, access is by application and top-ups are applied by our team on Telegram rather than through self-serve checkout.
Bonus credit — from a top-up bonus, an invite or our team — is spent before paid credit on every model except Model One, which is billed to paid credits only. Top-ups are final — we do not refund or pay out credit, paid or bonus — and credits do not expire monthly. The full rules are in the Terms of Service.
Yes. Select “Company” in the application and describe your requirements — volume, compliance questions, deployment constraints. Dedicated deployments on reserved capacity are on our roadmap and we're actively talking to companies about them.
During the beta, requests are forwarded to external inference partners to be processed. We collect what we need to authenticate, bill and operate the service, and we don't sell customer data.
We're not making zero-logging or isolation claims we can't yet back up. Optional PII detection and redaction before requests reach providers is on the roadmap. The details are in the Privacy Policy.
It is open now, by invitation. Applications are reviewed in order and access is granted in small batches, so there is no fixed waiting time. Apply below and we'll contact you through the channel you provide.
Private beta access is granted in small batches. Tell us what you're building — it takes two minutes.