NGARi Inference API

Bison Venture Partners LLC dba NGARi · v1 · Zero retention · Attested

The API serves qwen3.8-orin — a 27.3B sovereign model (Apache 2.0) running on our own edge hardware (NVIDIA Jetson AGX Orin) with no third-party model provider in the path. We store nothing. No prompt or response is ever written to disk. Every request is metered against your prepaid balance and recorded as a signed attestation you can verify.

Get an API key

Enter the email you will use for billing. We show your key once; it is stored only in this browser.

Models

Model	Role	Notes
`qwen3.8-orin`	Default — preview + API	27.3B (Q4_K_M), Apache 2.0. Served on-device; no third-party model provider in the path.
`ngari-ft-distilled-v5`	Distilled line — not served publicly	1.5B, Apache 2.0, fine-tuned by NGARi. Retired from the public path: asked to ground on a long facts block it confabulated details about our own programs, so the 27B is served instead.

The served model is pinned per deployment. GET /api/health reports the live model tag, whether it is resident, and the memory gate — so you can always verify what you are calling rather than trust a page.

Model weights for the NGARi-distilled line are published Apache 2.0 at huggingface.co/NGARiAI. The hosted API runs qwen3.8-orin on-device; to run the same model on hardware you own, see NGARi+ hardware.

Endpoint

POST https://ngari.ai/api/chat

OpenAI-compatible endpoint (self-hosted)

On a NGARi device the sovereign model backend also speaks the OpenAI wire format at http://127.0.0.1:11438/v1, so any OpenAI-compatible harness — OpenClaw, Hermes Agent, DeepSeek Harness, the OpenAI SDK, opencode — can run on local, on-device intelligence.

POST http://127.0.0.1:11438/v1/chat/completions
GET  http://127.0.0.1:11438/v1/models

Point your client’s base URL at it. Loopback clients need no key; a ngari_ Bearer key is required for any non-local client and meters usage exactly like the public API. Completed requests return a signed attestation receipt (header X-NGARi-Attestation, or an SSE comment on streams) and appear under GET /v1/attestations. This is the layer a plain open-source harness cannot provide: on-device inference plus a verifiable proof of where it ran.

Authentication

Pass your key as a Bearer token. Keys start with ngari_ and are shown only once at creation. We store only a hash — lost keys cannot be recovered (create a new one).

Authorization: Bearer ngari_your_key_here

Example

curl https://ngari.ai/api/chat \
  -H "Authorization: Bearer ngari_..." \
  -H "Content-Type: application/json" \
  -d '{"messages":[{"role":"user","content":"Explain sovereign AI in two sentences."}]}'

The response is an SSE stream: a first event carries {"model":...}, then {"delta":"..."} chunks, then [DONE].

Key management

Operation	Endpoint
Create key (email = billing identity)	`POST /api/keys` `{"email":"you@..."}`
Balance + usage	`GET /api/keys/me` (Bearer)
Signed attestation receipts	`GET /api/keys/attestations` (Bearer)

Pricing

Pack	Credits	Tokens	Link
$1	500	500K	buy.stripe.com/cNi4gz45M4tUc35d1ubMQ1J
$5	2,500	2.5M	buy.stripe.com/28EaEXfOu1hI7MP0eIbMQ1z
$20	10,000	10M	buy.stripe.com/cNiaEX31If8yffhd1ubMQ1A
$100	50,000	50M	buy.stripe.com/6oU5kD6dUbWm5EH9PibMQ1B

1 credit = 1,000 tokens · $2.00 per 1M tokens. Credits are charged on actual usage (an up-front budget is reserved per request and the unused remainder refunded automatically). Credits never expire. Unused credits are refundable within 14 days of purchase — email [email protected].

Purchases settle to your key by email: buy credits with the same email you used to create your key, and the balance appears within ~5 minutes.

Rate limits

300 requests/hour per key · output capped at 900 tokens per request
Without a key, the public preview is free and limited (40/hr) — the API tier is for programmatic use

The attestation receipt

Each paid request writes an HMAC-signed record to the NGARi sovereignty meter: model, timestamp, device ID, mode (on-device), and token count. GET /api/keys/attestations returns your recent receipts. This is the proof that your inference ran on sovereign edge hardware and that nothing was retained.

What this API is not

It is a capped beta on a single edge device — no SLA. It is not air-gapped end-to-end (traffic transits Cloudflare), it is not end-to-end encrypted, and it cannot claim "your data never touches our servers" (inference runs here). What it does guarantee, contractually and architecturally: zero retention of prompts and responses, and verifiable receipts. For true on-premise sovereignty, the model ships on the NGARi ONE.

Acceptable use

No unlawful activity, no resale of raw inference as a competing service, no reverse-engineering of the model. See the Terms.

NGARi+ hardware — run this model on a machine you own, with no network in the loop.
Agents — hire the same intelligence as six executives you reach by email.
Environment — measured power draw for this device and the method behind it.