Inference Providers#

MEGA Inference Providers is a routed, OpenAI-compatible inference service at inference.tensorplay.cn. Use one MEGA token to call models across supported Providers, choose a routing strategy, or use your own Provider key. Availability, performance, and pricing are live route properties, so start integrations from the current model catalog.

Routed calls include monthly credit: Community users receive $0.10, PRO users receive $2.00, and Team or Enterprise organizations receive $2.00 per billed seat. See Pricing and Billing for eligibility, settlement, and spending limits.

Choose a path#

Public endpoints#

The initial public release does not expose image, video, Gemini-native, or Anthropic-native APIs.

The public model comparison page expands every validated model into its live Provider routes and exposes the same price, context, capability, latency, and throughput fields as the catalog API. fastest and cheapest badges are computed within each model and task.

Authentication#

Send a MEGA PAT with inference:run:

Authorization: Bearer mega_...

Use a MEGA PAT as the bearer token. Provider credentials are configured separately and are never accepted as public MEGA bearer tokens. Token and membership changes can take up to 60 seconds to reach every inference request.

Model and Provider selection#

Model IDs use owner/model[:selection]:

Within one Provider, MEGA ranks healthy endpoints using available capacity, error rate, throughput, and recent latency. A route temporarily disappears when it is unavailable or fails compatibility checks. See Routing and Provider Selection for the complete eligibility and failover rules.

Only routes that pass MEGA's protocol, latency, and advertised-capability checks appear as live. The default compatibility standard requires Chat Completions and Responses to produce a first token within five seconds and Embeddings to finish within 30 seconds. Failed routes are removed from selection until they pass again.

Task-specific request examples are available for Chat Completions, Responses, and Embeddings.

Automatic key selection and Routed Inference#

With no billing header, the Router uses auto:

X-Mega-Inference-Billing: auto

For each selected Provider, the Hub first uses an eligible personal or organization custom key when one is configured. If no key is available, the same request falls back to Routed Inference. This matches the Settings page: adding or rotating a key changes future automatic calls without changing application code.

Force MEGA billing for one request with X-Mega-Inference-Billing: routed. Force a saved key, with no routed fallback, with X-Mega-Inference-Billing: byok.

The Hub preauthorizes a bounded maximum cost before dispatch. After completion, it reconciles the Provider's usage and standard price, settles the actual integer nano-USD amount, and releases the unused reservation. MEGA applies no additional inference markup.

Some Providers report final usage asynchronously. If MEGA cannot confirm the final cost within approximately 30 minutes, it releases the reservation and does not charge the account for that request.

Custom Provider Key (BYOK)#

Save a personal key or an organization key under Settings → Inference Providers. Keys are encrypted at rest and excluded from logs, analytics, catalog responses, and error messages.

Require BYOK per request:

X-Mega-Inference-Billing: byok

MEGA still authenticates the PAT and applies account preferences and permissions before dispatch. The same key selection happens automatically when the header is omitted and an eligible saved key exists. MEGA records route and token metrics but charges zero MEGA inference cost; the Provider bills the key owner. BYOK is available only for Providers that support custom keys.

Organization billing#

Set the organization handle explicitly:

X-Mega-Bill-To: research-lab

The organization must be on Team or Enterprise, and the caller must have write or admin membership. Organization Provider order, disabled list, custom keys, and monthly hard limit take precedence for that request. Without the header, the personal account is the billing owner.

Streaming and failover#

Chat Completions and Responses preserve Provider Server-Sent Events. The Router may try another healthy candidate only before response headers or the first stream byte are exposed. Once output begins, the request is never transparently replayed.

All responses include:

Inference-Id: inf_uuid

Any response dispatched to a Provider also exposes the selected Provider, including BYOK and upstream error responses:

X-Mega-Provider: provider-slug

Browsers may read both headers through CORS. Keep Inference-Id in application logs; do not add prompt content to an incident report.

Sticky Provider sessions#

Some OAuth-backed upstream accounts require conversation affinity. Pass a stable, non-secret session signal in the dedicated header:

X-Mega-Session-Id: conversation-42

The Router hashes this header together with the billing owner and model, then forwards only an irreversible mega_… session identifier. It can also derive affinity from previous_response_id, session_id, conversation_id, or prompt_cache_key in a compatible JSON body; those body fields remain part of the upstream protocol and are forwarded unchanged. Use X-Mega-Session-Id when the original application signal must not reach the Provider.

Stickiness does not override health, a disabled Provider, capacity admission, or an explicit Provider suffix.

Privacy and observability#

MEGA does not retain prompts, responses, embedding inputs, or tool arguments. It retains only content-free request metadata needed for routing, reliability, abuse prevention, and billing, such as request hashes, byte and token counts, model, Provider, cost, latency, and status. Authorization headers, cookies, and Provider keys are redacted. The selected Provider necessarily receives the content required to perform inference; review that Provider's data policy before use.

API quick reference#

1curl https://inference.tensorplay.cn/v1/responses \2  -H "Authorization: Bearer $MEGA_TOKEN" \3  -H "Content-Type: application/json" \4  --data '{5    "model": "OWNER/MODEL:preferred",6    "input": "Hello"7  }'

Shared request controls are:

For the runnable setup path, start with Your First Inference Provider Call.