Xiaomi MiMo text and streaming pilot

This local source pilot selects Xiaomi MiMo's usage-priced mimo-v2.6-flash model through https://api.xiaomimimo.com/v1/chat/completions. Its declared scope is text generation and streaming. Source preparation, public contract acceptance, installed-package acceptance, current account/API availability and delivery are separate gates. This draft gives no live, delivered-provider or provider-target credit.

Exact source facts

All pages below are on the vendor's own documentation host, mimo.mi.com, and were retrieved on 2026-10-10 (hashes at the end).

• Context and output integers. The model table (model.md) shows only the rounded labels 1M and 128K, so nothing was expanded from it. The integers come from two vendor-authored client configurations that bind them to the exact id mimo-v2.6-flash: the OpenClaw guide ("contextWindow": 1048576, "maxTokens": 131072) and the OpenCode guide ("limit": { "context": 1048576, "output": 131072 }). Both name https://api.xiaomimimo.com/v1 as the pay-as-you-go base URL. The two pages use different client schemas and agree. The API reference independently gives [1, 131072] as the required range of max_completion_tokens and a default of 131072 for mimo-v2.6-flash; that parameter covers visible plus reasoning tokens. No first-party prose or spec table states the context integer; the OpenClaw and OpenCode snippets are the only bindings, and 1048576 equals 1024 * 1024.
• Route and auth. The API reference documents https://api.xiaomimimo.com/v1/chat/completions and two alternative headers, api-key: $MIMO_API_KEY and Authorization: Bearer $MIMO_API_KEY. This entry uses the Bearer form.
• Lifecycle. mimo-v2.6-flash is current and is the named replacement for the retiring mimo-v2.5; it is not itself on a retirement schedule. The deprecation page gives mimo-v2.5 and mimo-v2.5-pro a 10:00 (UTC+8) switch on 2026-10-14 in its table and 18:00 (UTC+8) in a note on the same page. The discrepancy is preserved here; no exact switch instant is certified.

Billing

The API is usage-priced: the pay-as-you-go page says it "consumes the account balance based on the actual Token usage". The promotions page describes bonus credit, obtained through Refer & Earn or new-user registration, that "can be used to offset pay-as-you-go API usage" and is "currently valid for 40 days". It also says eligibility, credit amounts and payout rules vary by campaign. The catalog therefore uses promotional-credit, following the NAVER CLOVA Studio profile (src/lib/providers/catalog/naver-clova-studio.json). This is not a free tier. Whether a payment method or identity verification is required to obtain a working key was not established for the real-time API, the credit amount and eligibility were not checked on an account, and prices are not recorded (the pay-as-you-go page lists CNY and USD and they were not converted).

Scope that is deliberately not declared

The capability booleans limit the pilot's declared scope; false is not evidence that the vendor lacks a feature. Tools, vision, structured output and SDK thinking controls are not declared because none has been probed. The deprecation page notes that in thinking mode mimo-v2.6-flash ignores custom temperature and top_p, so a thinking mode exists; this entry exposes no control for it. Token Plan keys and base URLs are a separate subscription product and are out of scope.

Known gaps the local contract cannot close

• Output-budget field name. The vendor's request reference lists only max_completion_tokens. The SDK's shared chat-completions base sends max_tokens, and the suite pins that observed field. Whether the vendor honors max_tokens, ignores it (leaving the model's 131072 default) or rejects it is unverified. A catalog entry cannot rename the field; that would need a provider hook or a new catalog quirk, which is an operator decision.
• Streaming usage option. A streaming request carries stream_options: { "include_usage": true }. The vendor's reference documents stream but not stream_options. The suite does not observe how the vendor responds to the extra field.
• Environment variable. The vendor's own examples name MIMO_API_KEY. The catalog's derived primary variable is XIAOMI_MIMO_API_KEY, and MIMO_API_KEY is read as a fallback through wire.apiKeyFallbackEnvVars, the mechanism the Sakana AI entry also uses. Removing the fallback is a one-line change if the maintainers prefer the pure convention.
• Ceilings are source-bound only. The shared wire builder fits max_tokens only to runtime-discovered ceilings, so the catalog's contextWindow and maxOutputTokens never appear on the wire, and no public CLI command or SDK export was found that reads them back. The owned endpoint cannot observe them; they rest on the cited source bindings alone.

Local public contract

test/continuous-test-suite-xiaomi-mimo-pilot.ts runs 12 cases through the built public SDK and built CLI, with optional --package-root support for a later installed consumer:

Cases	Count	Required observation
SDK/CLI generate/stream default	4	Actual catalog default, route, Bearer header, unique user prompt, max_tokens: 32, stream mode, no tools/tool choice, and fresh answer
SDK/CLI generate/stream invalid explicit id	4	All wire preconditions pass before HTTP 400 rejection; no fallback or successful reply
SDK and CLI generate, only MIMO_API_KEY	2	The vendor-documented variable alone reaches the wire as the Bearer credential
SDK generate, both variables set	1	XIAOMI_MIMO_API_KEY wins; the fallback's value never reaches the wire
SDK generate, no key	1	Refusal before any request, naming XIAOMI_MIMO_API_KEY

The fixture strips real credentials before imports, uses temporary homes and sets the provider route to its owned random-port listener. Public MODEL_CONFIG_URL uses that same listener's HEAD /health and GET /api/v1/models. A schema-valid empty registry keeps the provider's actual catalog default in control. Positive metadata initialization requires the exact ordered request pair; explicit CLI model paths require no metadata request. CLI commands keep a 60-second deadline. The suite makes no vendor call, creates no account and reads no secret.

The validated source invocation and public-suite command are:

pnpm exec tsx tools/verify-provider-onboarding.ts --catalog-stage source --provider xiaomi-mimo
pnpm exec tsx test/continuous-test-suite-xiaomi-mimo-pilot.ts

Provenance and remaining gates

The legacy evidence.rosterVerified slot carries the explicit PUBLIC_DOCS_ONLY_NOT_ROSTER_VERIFIED marker and documentation methods; it has no HTTP status, auth probe or roster. billingProbe is likewise documentation only. liveMatrix is null. addedInPR holds the canonical introduction PR URL, https://github.com/juspay/neurolink/pull/1977; it was PENDING_PR in the first commit, before the PR existed. A URL alone is not proof of merge, and source-stage admission gives no provider-count, roster, live, installed-package or publication credit.

Authenticated roster, account creation, funding, live inference, installed consumer acceptance and delivery are separate and not done here.

Source receipts

Retrieved with a plain GET on 2026-10-10; the SHA-256 is of each response body.

First-party source	Body SHA256
Model table	26a7f23c7a12410e10cc75610dc9e6c364a637495161f5d41bc2a21e1075ed84
OpenClaw integration	4fab15ca323ac46f8c9905d5a82d7cb02023220f17d20ba21c69dce5b540c924
OpenCode integration	d39cb672b9f9da03f3da00e83c964849bfdf3f9d06ca54efbbabd5374aa8c376
OpenAI-compatible API reference	1be241b9a882b3c70954e8ec04031cb657b06c8e0b0bbb8511e5b6ec63b82ae7
Promotions FAQ	63e8515066b5ec06a61e71262b4a5f41b945f9bd4e540b76299f3437880223e7
Pay-as-you-go pricing	f62427a339e91edbb252cc6b54cac92050d4254483b1e7eb97a234d6d32f947b
Deprecation schedule	f5c856884116df8ebba552550ec5c6912b1d0bd8edad66e2bf1e973da6270068