pi-gmicloud-ai

Pi extension for GMI Cloud Inference Engine (IE) — the hosted inference platform behind console.gmicloud.ai.

It wires GMI Cloud's OpenAI-compatible API into Pi so you can use the whole GMI model library (DeepSeek, GPT, Claude, Gemini, Llama, Kimi, GLM, Qwen, …) directly from Pi.

Setup

export GMI_API_KEY=eyJhbGci...          # JWT from https://console.gmicloud.ai → API Keys

Or store the token through Pi:

/login gmi

Pi saves it in ~/.pi/agent/auth.json; GMI_API_KEY remains supported as a fallback. This also applies to /gmi-tts, /gmi-music, /gmi-image, and the model-listing commands.

Install like any pi extension:

npm install pi-gmicloud-ai

Then restart pi so the gmi provider is loaded.

Capabilities

Capability	Endpoint	Status
Chat (text)	POST https://api.gmi-serving.com/v1/chat/completions	✅ live-verified (base URL from official docs)
Vision (VLM, image→text)	same, image_url content part	✅ OpenAI-compatible (delegated to pi core)
Embeddings	POST https://api.gmi-serving.com/v1/embeddings	⚠️ catalog-flagged, not streamed through this path yet
Speech (TTS)	POST /api/v1/ie/requestqueue/apikey/requests (poll + download)	✅ wired as /gmi-tts (MiniMax Speech 2.8)
Music	POST /api/v1/apikey/requests (poll + download)	✅ wired as /gmi-music (MiniMax Music 3.0)
Image (text→image)	POST /api/v1/ie/requestqueue/apikey/requests (poll + download)	✅ wired as /gmi-image (e.g. seedream-5.0-pro) — not free (402 without balance)

API base

GMI Cloud exposes a single OpenAI-compatible base — no second "genai" host and no special-cased request body (unlike NVIDIA NIM):

• https://api.gmi-serving.com/v1 serves chat, vision, and embeddings (addressed by model in the request body). This is what the extension uses.

You can see the exact base URL in GMI's own docs — e.g. the CURSOR integration guide tells you to set the "Override OpenAI Base URL" to https://api.gmi-serving.com/v1.

Model IDs

GMI model IDs use the provider/Model-Name form, e.g.

• deepseek-ai/DeepSeek-R1-Distill-Llama-70B
• openai/gpt-5.2
• anthropic/claude-opus-4.7
• meta-llama/Llama-4-Maverick-17B-128E-Instruct-FP8

The full library is listed at docs.gmicloud.ai/model-quickstarts/model-library.

Media generation commands (TTS / Music / Image)

GMI's audio/music/image models use a separate async request-queue API on the console host (not OpenAI-compatible), so they are exposed as pi commands rather than chat models:

• /gmi-models [chat|audio|music|image|video|all] [--free|--all] — list GMI models with a 🟢 FREE badge and capability. Chat models show Chat / Reasoning / Vision / Embedding; media models show TTS / Voice Clone / Music / Image / Video. Chat + media default to FREE-only (chat: MiniMax M3 / M2.7; media: Speech 2.8 TTS / Voice-Clone + Music 3.0); pass --all to list every model, or a capability like image --all to expand just that group. The FREE badge is data-driven: chat FREE = the live /v1/models pricing is 0 (GMI lists promo models as a duplicate zero-price row — verified for MiniMax-M3 / MiniMax-M2.7); media FREE = verified promo families (Speech 2.8, Music 3.0). Pass --free to list only free models.
• /gmi-tts <text> — MiniMax Speech 2.8 TTS. Options:
  • --voice <id> — voice id (default English_expressive_narrator)
  • --model minimax-tts-speech-2.8-turbo|hd — model (default …-turbo)
  • Example: /gmi-tts --voice English_expressive_narrator Hello from Pi
• /gmi-music --prompt "<style>" <lyrics> — MiniMax Music 3.0. Example:
  • /gmi-music --prompt "lo-fi, chill" [verse]\nWalking down the empty street
• /gmi-image [--model <id>] [--size 2K|3K|WxH] <prompt> — text→image (default seedream-5.0-pro). Example:
  • /gmi-image a watercolor lighthouse at sunset
• /gmi-audio-models [tts|music|all] — list the available audio/music model ids, display names, and descriptions (handy for picking a --model/--voice).

TTS and music POST a generation request, poll until status: success, download the resulting audio, save it to .pi/generated-audio/, and show a clickable file link (pi has no inline Audio renderer, so the file opens in your player). Images render inline in the TUI via pi-tui's Image renderer and are also saved to .pi/generated-images/.

Requires GMI_API_KEY. During the 2026-08-24 → 09-06 promo, Speech 2.8 and Music 3.0 are free (H3 is excluded). Image generation is NOT in the free range and returns 402 Insufficient balance on a zero-balance account.

Free promo (2026-08-24 → 2026-09-06)

GMI Cloud + MiniMax are running a 14-day free campaign. On a zero-balance account (no credits purchased), only these are verified free during the promo:

• MiniMaxAI/MiniMax-M3 ✅ (chat — verified 200 through this extension)
• MiniMaxAI/MiniMax-M2.7 ✅ (chat/reasoning — verified 200 through this extension)
• MiniMax Speech 2.8 and Music 3.0 (audio / music — wired as /gmi-tts and /gmi-music; TTS verified end-to-end)

💡 To chat for free right now, select MiniMaxAI/MiniMax-M3 or MiniMaxAI/MiniMax-M2.7 as your model. Any other chat model (e.g. zai-org/GLM-5.3-Flash, openai/gpt-*, deepseek-ai/*, meta-llama/*, …) returns 402 Insufficient balance on a zero-balance account — that is expected, not a bug. Add credits to use the rest of the 80+ model catalog.

⚠️ Other MiniMax chat models (MiniMax-M2.5, MiniMax-M1, etc.) were not verified free and also returned 402 — stick to M3 / M2.7 for free chat. ⚠️ MiniMax-H3 is NOT in the free range — it still requires account balance. ⚠️ Image generation is also not in the free range — /gmi-image (e.g. seedream-5.0-pro) returns 402 Insufficient balance on a zero-balance account. The full catalog is still fetched from /v1/models; just pick a free id during the promo to avoid the 402.

Notes

• The seed list is a representative subset of IDs taken from the live /v1/models response (so instant startup never offers a stale id). The full catalog (80+ models) is fetched from /v1/models in the background and cached on disk. The live catalog is broader and newer than the docs model library snapshot — e.g. it includes openai/gpt-5.5, google/gemini-3.7-flash, anthropic/claude-opus-4.7, moonshotai/kimi-k3, x-ai/grok-4.5, Qwen/Qwen3.8-*, and others.
• Each model's context_length and per-token pricing (prompt / completion / cache read / cache write) are read from /v1/models and converted to Pi's per-million-token cost fields, so /list-models and usage accounting reflect real GMI prices.
• Reasoning models (*-r1, *-Thinking, o1/o3, qwq, …) are flagged via id pattern so Pi can route thinking budgets correctly.
• Vision (image-input) capability is detected from the live /v1/models response when GMI reports modalities; otherwise models default to text input.
• Embedding models in the library (e.g. laion/CLIP-ViT-B-32-...) are flagged but not streamed through chat completions in this extension yet.