# Omnia Model Hub — full text Omnia (Omnia Model Hub) is a neutral marketplace for model inference API calls. It gives developers one OpenAI-compatible endpoint, lets suppliers publish their own quotes for the same model, verifies supply with probes and five-level trust grades, and charges a buyer-side transaction fee on each successful call. ## Omnia Model Hub — One API, every latest model Call the newest frontier models from OpenAI, Anthropic, Google, xAI and more through one OpenAI-compatible endpoint. Compare suppliers, prices and trust grades before you call. ### What Omnia is Omnia is a neutral marketplace for model inference. One endpoint, every model: developers bring their own key and route among suppliers that quote the same model. Omnia is the intermediary, not the seller. - One OpenAI-compatible base URL for every model. - The same model is quoted by several suppliers, so price and latency are side by side. - Probes and five-level trust grades make supply verifiable, not promised. ## Features — Omnia Model Hub All fifteen modules are developed as one release, not in phases. ### Gateway and protocol compatibility One base URL, OpenAI-compatible request shapes, API key authentication, rate limits and budgets. Provider routing objects travel with every endpoint. ### Marketplace and quoting A resource is supplier × model × region × price × service-level agreement × service tier. The same model can have several quotes. Model version pinning, change notices and migration hints are recorded. ### Routing and reliability Six production signals — effective price, throughput, first-token latency, 429 rate, failure rate and tool-call quality — are combined with trust grade into a deterministic weighted score. Every decision records which resources were considered and why one was chosen. ### Metering, billing and credits Billing rules are explicit, pure and unit-tested: cached input is priced separately, failed retries are ignored, empty completions produce a zero bill, and the platform fee is computed on the subtotal. Bills have line items and can be recomputed. ### Verification and trust Admission compares model fingerprints and runs capability, performance, availability and capacity tests. Listings are then sampled continuously. Results become a five-level grade; a downgrade reduces routing priority, limits traffic or suspends the listing, with an append-only change history. ### Data governance, teams and analytics Zero retention is the default. Three data policy tiers, region-aware routing, guardrail rules, organisation workspaces, budgets with 80 percent alerts and 100 percent stops, activity drill-down, public rankings with a published methodology and app attribution are included. ### Administration Administrators use a separate sign-in page and a separate account store. Every function is controllable from the back office, including the transaction fee rate, user status and permissions, listings, ratings, invites, disputes, content, feature flags and configuration. ## Model library — Omnia Model Hub Omnia aggregates the world's frontier models behind one OpenAI-compatible endpoint. Browse the catalogue; live calls and pricing open as suppliers onboard. ### One endpoint for the whole catalogue Point any OpenAI-compatible client at the Omnia base URL, create an API key in the console, and call any model on this page. The same model may be quoted by several suppliers, so price and latency are compared for you. ### Live availability Cards marked "Live" have active supplier listings you can call right now. Cards marked "Suppliers onboarding" are in the catalogue and become callable as suppliers list them — no action needed from you. ## Pricing — Omnia Model Hub Suppliers set quotes. Omnia charges one transparent buyer-side fee on each successful call. ### Fee structure The current platform fee is 5 percent of the call subtotal. The rate is stored as a versioned, append-only setting and can be changed by an authorised administrator; a change applies to the next settled call only. - Input, cached input and output tokens are priced separately per million tokens. - Failed retries and empty completions are not billed. - Web search, file parsing and media auxiliary services are separate bill lines. - Bills are recomputable from usage events and quote snapshots. ### Plans Four tiers — Free, Standard, Business and Enterprise — control rate limits and feature gates. The Enterprise tier adds SSO, SCIM, workload identity federation, DPA, contractual SLA and audit export. ### Credits and payments Service credit is prepaid service fee and can only be consumed on the platform. Invite credit can only offset model-call bills. Neither balance is withdrawable, transferable or redeemable. Payments are not connected in the development build; an isolated adapter is added at launch. ## API documentation — Omnia Model Hub Point an OpenAI-compatible client at the Omnia base URL and use your platform API key. ### Authentication Send Authorization: Bearer . Keys are shown once at creation and stored only as a SHA-256 hash. Keys can be revoked from the console or by an administrator. ### Endpoints POST /v1/chat/completions, POST /v1/responses, POST /v1/completions, POST /v1/embeddings, POST /v1/images/generations, POST /v1/video/generations, POST /v1/audio/transcriptions, POST /v1/batches and GET /v1/models. ### Provider routing object The provider object supports order, only, ignore, allow_fallbacks, sort, data_collection, zdr, require_parameters, quantizations, max_price, preferred_min_throughput, preferred_max_latency, enforce_distillable_text and ignore_performance. Unknown fields are rejected rather than ignored. ### Example request POST /v1/chat/completions with model, messages and an optional provider object. The response contains standard choices and usage fields plus an omnia object with the chosen resource, fee rate, subtotal, platform fee, auxiliary charge, total and route reason. ### Recomputable bills Every response includes a bill identifier. GET /api/user/bills/ returns the bill header, all line items and the settlement split between invite credit and service credit. ## Frequently asked questions — Omnia Model Hub Direct answers, written so they can be quoted accurately. ### What is Omnia? Omnia (Omnia Model Hub) is a neutral marketplace for model inference API calls. It gives developers one OpenAI-compatible endpoint, lets suppliers publish their own quotes for the same model, verifies supply with probes and five-level trust grades, and charges a buyer-side transaction fee on each successful call. ### Does Omnia sell model weights or tradable tokens? No. Omnia trades inference API service. It does not sell model weights, does not offer secondary transfer of tokens, does not run an order book, and does not allow shared API keys. ### Is the API compatible with OpenAI clients? Yes. Omnia exposes OpenAI-compatible request and response shapes, so existing clients can change the base URL and use a platform API key. ### How is a call billed? Input, cached input and output tokens are metered separately at the supplier's per-million-token quote. A 5 percent buyer-side transaction fee is applied to the subtotal. Failed retries and empty completions are not billed. ### Can I withdraw my balance or invite credit? No. Neither service credit nor invite credit is withdrawable, transferable or redeemable for cash. Invite credit can only offset Omnia model-call bills. ### How does the invite programme work? Every account automatically receives one unique invite link. A qualifying first top-up by a referred user creates a fixed platform credit for the inviter after a cooling-off period. There is one level only, no downline, no recruitment reward and no cash payout. The material connection must be disclosed when sharing the link. ### Who controls the transaction fee? An authorised administrator changes the rate in the separate administration console. Each change is versioned, append-only and audited, and applies only to later calls. ### Are administrator accounts stored with user accounts? No. Administrators use a separate sign-in page, a separate account table, a separate session cookie and a mandatory authenticator code. ### Which languages does the interface support? English is primary and Simplified Chinese is available on every page and screen. A globe icon in the top right switches the language. ### Is Omnia a securities or crypto exchange? No. The name Omnia deliberately avoids regulated exchange wording. The platform does not provide securities trading, crypto-asset trading, stored-value redemption or cash payouts. ## Security — Omnia Model Hub Security controls are implemented in the application, not promised as a future phase. ### Identity separation Administrators and users have separate account tables, separate sign-in pages, separate session cookies with separate namespaces and separate session lifetimes. Administrator sign-in requires a six-digit authenticator code; prior admin sessions are revoked at each sign-in and sessions are bound to the originating IP. ### Application controls Passwords use scrypt with a per-account salt. Sessions use opaque random tokens stored as hashes. Every state-changing authenticated request requires a per-session CSRF token. Login endpoints have per-account and per-IP failure lockouts; API and gateway endpoints are rate-limited; blocked IPs are rejected before processing. ### Data protection All SQL uses bound parameters with column allow-lists, so a client cannot inject table or column names. Sensitive provider keys are encrypted at rest with authenticated encryption and a file-based master key. API keys are stored as hashes and shown once. JSON responses are rendered with text nodes on the client, and security headers include a restrictive Content Security Policy, frame denial, nosniff and no-store on APIs. ### Integrity and audit Bills, usage events, both credit ledgers, settlements, fee versions, probe records, rating changes, route decisions, security events and audit logs are append-only at the database level through SQLite triggers. Administrator actions record actor, target, before and after values, and IP. ### What deployment still adds Production deployment should add TLS termination at a reverse proxy, host firewall, encrypted backups, secret rotation, log shipping and a Web Application Firewall. These are infrastructure tasks, not missing application features. ## About Omnia Name, scope and operating sequence. ### Name The global brand is Omnia, pronounced OM-nee-uh, from the Latin word for all things. The Chinese name is 万象. The product descriptor is Omnia Model Hub. The name deliberately avoids regulated exchange-style wording. ### Operating sequence The system is built and tested first. Payment is connected next. A Hong Kong entity and trademark registration are considered only after the platform has operated and generated revenue. ### Language and region English is the primary language and Simplified Chinese is a complete secondary language. The platform is intended for global users; mainland China is not the initial target market. ## Status — Omnia Model Hub Service state is read from the platform's own incident records. ### Current state The public status endpoint reports operational unless an open incident exists. Administrators create and resolve incidents in the back office. ### Observability Route decisions, probe records, security events and audit logs are retained with timestamps. Bills and usage events are append-only so historical figures cannot be silently rewritten. ## Changelog — Omnia Model Hub Dated changes, written for people and for retrieval systems. ### 2026-11-15 · Version 1.0 First complete release. All fifteen modules were developed in one pass: gateway, marketplace, routing, metering, credits, verification, data governance, teams, analytics, plans, supplier side, localisation, developer platform, automatic invite credit and operations. - Global brand set to Omnia with 万象 as the Chinese name. - English primary and Simplified Chinese complete, with a top-right language switch. - Separate administration console with full permission control over every function, including the transaction fee. - Payment adapter intentionally disabled until launch; offline credit supports end-to-end testing. - Official website published with structured data, llms.txt, robots.txt and sitemap.xml. ## For suppliers — publish your own model capacity on Omnia Apply for an account, review is done by an administrator, and after approval you manage your own upstream credentials, quotes and listings. Omnia is the intermediary, not the seller. ### How the supplier side works A supplier submits one application with its company name, contact, email and a self-chosen password. An administrator reviews the application. Approval creates a supplier account, a supplier profile and an initial verification record; the supplier then signs in at the supplier console with the password set at application time. The supplier publishes its own quotes for a model, region, price, latency, throughput and service level. Buyers see those quotes, Omnia routes among eligible suppliers, meters the call and produces a recomputable bill. A buyer-side transaction fee of 5 percent is charged on each successful call. - Administrators review applications, freeze accounts, set the platform fee and arbitrate disputes. - Administrators do not enter upstream addresses or keys for an onboarded supplier and cannot change its prices. - Payment is not connected in the development build; settlement reports are produced per supplier but no funds are cleared. ### What you control in the supplier console The upstream base URL, authentication scheme, encrypted API key, custom headers, timeout and retry policy are entered by the supplier itself. Keys are encrypted at rest with authenticated encryption and can be rotated or removed by the supplier at any time. When a call is routed, the platform forwards it using those credentials and never displays the stored value. Listings, per-million-token input / cached-input / output prices, region, model version, service tier and availability are all published and changed by the supplier. A listing can be suspended without deleting its history. ### Verification and trust records, bound to the supplier entity Admission checks compare model fingerprints and run capability, performance and availability tests. Listings are then sampled with scheduled probes. Fingerprint and probe records carry the supplier identifier, and the results are aggregated into one verification status and one five-level grade per supplier entity, not just per listing. Grade changes are written to an append-only history and reduce routing priority, limit traffic or suspend a listing. The same records are visible to the supplier that owns them. ### Bitcoin-anchored evidence Audit records are batched into a Merkle tree and exported as offline-verifiable anchor packages. The SHA-256 digest of each package is submitted to public OpenTimestamps calendars in three stages: stamp (pending calendar receipt), upgrade (complete Bitcoin attestation with a block height) and verify (independent re-check against Bitcoin). Only a 32-byte digest leaves the server; audit text and business data never go on a public chain. The proof establishes that the digest existed before a specific Bitcoin block. It does not prove that the content was true, and it is not a court ruling on admissibility. ### Do I have to send my upstream API key to an administrator? No. After approval you enter it yourself in the supplier console. It is encrypted at rest, never shown again, and can be rotated or removed by you. The platform holds the encrypted credential only to forward the calls you accepted. ### Can Omnia change my published prices? No. Quotes and listings belong to your supplier account. An administrator can freeze a supplier or suspend a listing for enforcement, but cannot edit your prices or upstream credentials. ### What does the Bitcoin timestamp actually prove? It proves that the hash of an anchor package existed before a specific Bitcoin block was mined. It does not prove the underlying business data is accurate, and a pending calendar receipt is not yet a completed proof.