xiaomi-mimo

> This is a discovery protocol, not a reference. It teaches you where to look for current MiMo capabilities — it does not mirror them. Anything API-shaped (model IDs, request schemas, voice catalogues, quotas) drifts faster than this manual; fetch the live docs every time.

Where to look

Two URLs and an LLM-friendly bulk dump are all the agent needs to bootstrap full knowledge of the API:

| Purpose | URL | |---|---| | Doc index (start here for any specific question) | platform.xiaomimimo.com/llms.txt | | Full doc dump (one curl → every public doc page concatenated) | platform.xiaomimimo.com/llms-full.txt | | Marketing / model gallery (skim once for the vibe) | mimo.xiaomi.com |

bash
# Quick orientation — what doc pages exist right now?
curl -s https://platform.xiaomimimo.com/llms.txt | head -60

# Full dump — when you need everything in context (much larger)
curl -s https://platform.xiaomimimo.com/llms-full.txt | wc -l

The doc index is structured so you can grep for what you need (e.g. grep -i tts, grep -i pricing, grep -i multimodal). Each entry has a verbatim URL — fetch the specific page once you've located it.

Roughly what's behind a MiMo API key

Enough context for the agent to know what kind of question to ask the docs. Verify model IDs and capability claims against the live docs every time — Xiaomi rotates suffixes and adds new variants regularly.

Three rough families share one https://<host>/v1/chat/completions endpoint:

• Text-only chat models — long-context reasoning and tool use. Use for plain LLM work; one is the 1M-context flagship.
• Multimodal-input chat models — accept image, audio, AND video as content parts (type: "image_url" | "input_audio" | "video_url") alongside text. Output is text. Use for transcription, image OCR, scene description, audio-visual joint reasoning, etc.
• Text-to-speech models — accept text, return base64-encoded audio. Three flavours: a built-in voice catalogue, a voice-design variant where you describe the voice in natural language, and a voice-clone variant where you upload a reference audio sample.

To get the current exact model IDs and per-model capability matrix:

bash
# Inspect the OpenAI-compat spec — lists every model the endpoint accepts
curl -s https://platform.xiaomimimo.com/static/docs/api/chat/openai-api.md | head -200

# Or hit the doc index and follow the API path it advertises
curl -s https://platform.xiaomimimo.com/llms.txt | grep -i 'api/chat'

For the modality you actually need, fetch the dedicated guide. The doc tree currently follows static/docs/usage-guide/<topic>.md — grep usage-guide against llms.txt to find the current path, then curl it.

Sourcing the API key

The TUI stores keys in ~/.lingtai-tui/.env and tells each preset which slot to read via manifest.llm.api_key_env. Slots are per-preset, so a user with both pay-as-you-go and a Token Plan account has two distinct env vars.

Resolution: scan presets, find MiMo ones, read their declared slot. Walk ~/.lingtai-tui/presets/*.json; for each manifest.llm.provider == "mimo", print its api_key_env slot (default MIMO_API_KEY if empty) and base_url so you can pick the right account/region. (For a recursive JSONC-safe scan, reuse the sibling ../minimax-cli/SKILL.md §3 scanner with provider == "mimo".)

Slot naming (per tui/internal/preset/preset.go::AutoEnvVarName):

• New user-saved presets: MIMO_<N>_API_KEY (no region suffix; N is a counter).
• Built-in / legacy presets: MIMO_API_KEY for back-compat.

Once you've picked the right slot, export it for any direct API call:

bash
SLOT=MIMO_1_API_KEY    # whichever slot the preset scan returned
export MIMO_API_KEY=$(grep -E "^${SLOT}=" ~/.lingtai-tui/.env | cut -d= -f2- | tr -d ' ')
export MIMO_BASE_URL=$(...)  # see "Picking the host" below

If multiple MiMo presets exist and you can't infer which the human means, ask — don't guess.

Picking the host

MiMo issues two key formats that pair with different host families. They are not interchangeable — using a tp- key against api.xiaomimimo.com or vice versa returns 401 invalid_key.

| Key prefix | Host family | Notes | |---|---|---| | sk-… | api.xiaomimimo.com (single global host) | Pay-as-you-go, per-token billing | | tp-… | token-plan-{cn,sgp,ams}.xiaomimimo.com | Token Plan, three regional clusters |

For Token Plan, the user's assigned cluster is shown on their Subscription page (platform.xiaomimimo.com/#/console/plan-manage) — the platform pins each plan to one cluster. Don't guess; ask the user, or read manifest.llm.base_url from the preset.

bash
# Note: bare URL is informational; real endpoint is /v1/chat/completions
# Inspect the key prefix
case "$MIMO_API_KEY" in
  sk-*) export MIMO_BASE_URL="https://api.xiaomimimo.com" ;;
  tp-*) echo "Token Plan — set MIMO_BASE_URL to the regional cluster from the user's preset" ;;
  *)    echo "unknown prefix — verify with the user" ;;
esac

For the Anthropic-compat surface (Token Plan only), the path is /anthropic instead of /v1. Both cluster types speak both APIs.

Switching models in the active preset

The default preset uses one of the multimodal-input models. To swap:

1. Run /setup, pick mimo, edit the manifest's llm.model field.
2. Or clone the preset (cp ~/.lingtai-tui/presets/mimo.json …/mimo-pro.json) and edit model.

Important: if you switch to a text-only model, also remove the vision capability from the manifest — otherwise the kernel's vision tool will fire against a text-only model and 400 on image input.

When the docs contradict this skill, file a report

This skill ages. The live docs are authoritative; if you find a discrepancy — a model that no longer exists, a doc URL that 404s, a request shape that returns a clean error against the documented spec, a new capability MiMo shipped that this skill doesn't mention — invoke the lingtai-issue-report skill to assemble a structured report. Don't silently work around the drift; surfacing it is how this skill stays accurate for the next agent. Specific signals worth reporting:

• llms.txt no longer returns the doc index (or moved)
• A model ID listed in static/docs/api/chat/openai-api.md returns "model not found"
• The cluster URL family in this skill's "Picking the host" table no longer matches the platform's current console
• A capability advertised at mimo.xiaomi.com (e.g. a new model variant) is missing from the API surface — or vice versa
• Pricing or tier limits in the docs differ from what usage fields actually report

Failure modes

| Symptom | Likely cause | Fix | |---|---|---| | 401 invalid_key | Wrong key prefix for the host (sk- vs tp-) | Match per the "Picking the host" table; lingtai-issue-report if the doc no longer matches the real prefixes | | 404 on chat completions | Token Plan account hitting api.xiaomimimo.com (or vice versa) | Swap MIMO_BASE_URL to the right family | | Token Plan latency spikes | Wrong regional cluster | Check the user's Subscription page; switch base_url to the assigned cluster | | 429 Rate limited | Hit RPM/TPM limit | Live docs — current per-model RPM/TPM caps live in static/docs/pricing.md | | Vision tool 400s after model swap | Switched to a text-only model but kept vision capability | Remove vision from manifest, or switch model back to a multimodal-input one | | Anything else weird | The skill is stale, the docs moved, or MiMo changed behaviour | Fetch llms.txt, follow the trail; if the trail itself is broken, file via lingtai-issue-report |

> Found a bug or issue? If you encounter any problems with this skill, load the lingtai-issue-report skill and follow its instructions to report it.