LooseMouth
Google Gemini handles hosted chat and research. Guest chat is available; saved conversations require an account.
Try guest chatBuild with evidence
Try it. Inspect it. Reproduce it.
Start with a small streaming request, then inspect the implementation and its limits. GIC's experimental checkpoints, Gemini-backed LooseMouth, and local BitVision inference are separate systems.
1. Make a streaming request
The public research API supports Facts-500 and SFT-1000. Examples use a short prompt and a low temperature for a small reproducible smoke test. A successful request demonstrates serving, not general model quality.
Use this model from cURL or Python
Calls GIC's public research endpoint. No token is included. Your prompt and output go through Vercel and Hugging Face. The same 8 requests per IP per hour and shared ZeroGPU quota apply.
curl -N 'https://www.globalintentcompany.space/api/intent/mini' -H 'Origin: https://www.globalintentcompany.space' -H 'Content-Type: application/json' --data-raw '{"message":"The capital of France is","checkpoint":"facts500","temperature":0.45,"topK":25,"maxTokens":32}'Responses are newline-delimited JSON. Each text field replaces the prior output. Check error events and the final done flag; HTTP 200 alone does not mean generation succeeded. Python uses only its standard library.
| Field | Supported values |
|---|---|
| checkpoint | facts500 or sft1000 |
| message | 1–8000 characters; the model has a 512-token serving context |
| temperature | 0–10; playground default 0.45 |
| topK | Integer 1–100; default 25 |
| maxTokens | Integer 1–128; playground default 96 |
| Limits | 8 requests per IP per hour, plus the shared HF account’s ZeroGPU quota |
Set the Origin header to the site origin shown in the example. The browser never receives the server's HF_TOKEN. There is no uptime or throughput guarantee. HTTP 400 indicates invalid input, 403 an origin mismatch, 429 a quota limit, and 502/503 an upstream or configuration issue. Stream errors may also arrive after HTTP 200.
For requests using your own Hugging Face quota, call the Space with your own token using the Space's Use via API instructions. Keep that token on your server. Authentication does not provide unlimited GPU time.
2. Know where computation happens
INTENT Mini
Vercel forwards each prompt to GIC’s custom checkpoint on Hugging Face ZeroGPU. Each generation is independent.
Try the research checkpointsBitVision Cell
The browser downloads ONNX Runtime and model weights, then performs inference locally via WebGPU or WASM. Local computation still requires initial network downloads; the surrounding site may use analytics.
Run the browser experimentVirtual Lab
Paid / Enhanced access is checked on the server. RHO and reference runs use bounded server solvers; synthetic optics and file inspection run in the browser. Optional AI uses Gemini.
Open the subscription cockpitEngineering note · October 16, 2026
From an anonymous embed to authenticated streaming
The first website integration embedded the public Space. Adding HF_TOKEN to Vercel could not authenticate requests originating inside that iframe. The native panel instead submits a validated prompt to a server route, which authenticates both Gradio job submission and SSE retrieval. The browser receives cumulative text snapshots as newline-delimited JSON.
A production test exposed a Gradio compatibility issue: a custom session identifier prevented completed jobs from being found by the result endpoint. Letting the /call API manage its own job session fixed retrieval. A short production smoke test then streamed “Paris.” for a capital-of-France prompt. This is evidence that serving worked for that request, not a benchmark of factual accuracy.
The regression suite checks token forwarding without disclosure, both checkpoints, fragmented Unicode streams, input and origin validation, persistent quotas, safe errors, incomplete streams, and cancellation. Stopping the web stream does not guarantee that a GPU task already running will stop.
- Inspect the server route
- Inspect and run the inference checks
node --test scripts/verify-intent-hf.cjs - Inspect the session compatibility fix
3. Share settings and inspect evidence
In the paid Lab, copy a RHO or optical settings link to reproduce the configuration. Links contain only bounded numeric settings, never access tokens, imported files, results, or approval. Recipients still need an active subscription. Restored settings do not run automatically.
Export RHO evidence, synthetic optical packets, or a model answer. Simulation output remains labeled SIMULATED. SHA-256 and hash links detect changes to the supplied bytes; unsigned exports do not establish author identity, hardware provenance, clinical validity, or an immutable record.