Global Intent Company · Foundation architecture research
Explore a 187.3M-parameter research language model trained from random initialization on a single NVIDIA L4 GPU. This checkpoint follows 10,000 pretraining steps, SFT-v2 step 750, then 500 facts fine-tuning updates.Its architecture combines 10 causal encoder blocks, four shared-weight recurrent controller blocks repeated across four search passes, three candidate branches per pass, 512 learned memory slots, a two-block auxiliary JEPA-style world model, and a latent critic.
187.3M parameters
4 passes · 3 candidates
Facts-500
Experimental model: responses can repeat, hallucinate, or produce incorrect code. The displayed critic values are uncalibrated internal scores, not correctness probabilities. The demo runs on Hugging Face ZeroGPU, not GIC's usual LooseMouth inference backend. Avoid entering private or sensitive information.
Authenticated Hugging Face connection · Adjustable decoding · Streaming output
Facts-500 · Streaming inference
Ready
Each prompt is an independent generation. Up to 8 requests per hour; shared Hugging Face GPU availability applies.
Stopping closes your stream. A GPU task already running may still finish on Hugging Face.
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":96}'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.
Inference runs on Hugging Face through GIC's authenticated server connection. Access here does not require a GIC subscription. The direct Hugging Face studio uses its own sign-in and quota.