Connect AI · C1

Independent brains. Connected intelligence.

Built by Jay · Connect AI LAB. We build AI in public and turn the process into practical education—from running local AI to building models of your own.

YouTube · CONNECT AI LAB · Connect AI LAB research · Learn with Jay · AI City Builders

A small, locally runnable decision model for Korean company workflows and simulation research. C1 scores a supplied list of actions from context. This first release contains the original MLX LoRA adapter used in our v0.1 experiments, plus a minimal local inference client.

Visual research page · Connect AI · Source and public comparison

Build AI with Jay

Go from using AI to building your own. Connect AI LAB shares the experiments behind our tools and models; AI City Builders connects the learning to AI-native solo businesses.

• Watch the build: Connect AI LAB on YouTube. Start with the original local-AI / second-brain series: EP.1, EP.2, EP.3, and the Connect AI v1 application release. These videos cover the original application journey, not a C1 training tutorial.
• Explore the research: Connect AI LAB — connected-model experiments and research materials, with member access.
• Learn and build: AI City Builders — Jay's education and community for AI-native solo businesses.

What is released

This repository contains the original custom MLX adapters. For Ollama and LM Studio, use the separately verified C1 Q8_0 GGUF package. Its company accuracy is 65.42%; the original MLX results below are unchanged. The GGUF package now includes Decision Studio: context + question + actions → real relative probabilities and the top action, calculated from token logprobs. Open the public web Decision Studio and connect your own LM Studio or Ollama API—no Python client required for the web interface. The GGUF card includes CORS setup. The existing c1_local.py --serve interface is an optional local alternative; ordinary free-form chat does not display these probabilities. This is a client update, not a new training run. The release is text-only decision inference; image/video use has not been evaluated. The original Connect AI v1 desktop application remains a separate product.

Run locally on an Apple silicon Mac

python3.12 -m venv .venv
source .venv/bin/activate
pip install huggingface-hub==1.33.0
hf download WonseokJayJung/Connect-C1-0.8B-v0.1 --local-dir Connect-C1
cd Connect-C1
pip install -r requirements.txt
python c1.py --input example.json

The first run downloads the pinned Qwen base model. Later inference runs locally using cached files. For a fully offline run, pass an already downloaded base directory:

HF_HUB_OFFLINE=1 python c1.py --model /path/to/Qwen3.5-0.8B --input example.json

example.json contains a fictional Korean request, a question, and options with unique key and description fields. The client selects the highest-scoring option token and returns all normalized option scores. These scores are not calibrated probabilities of correctness. It does not execute actions or implement automatic human handoff.

Recorded evaluation

Original adapter, default rule (highest option score), no policy text prepended:

The downloadable adapter was re-run on all 360 internal cases on 2026-10-02: 360/360 predictions matched the stored original run. validation.json records this check. The internal evaluation dataset is not released here, so that check is not independently reproducible from this repository alone. Training/evaluation overlap has not been independently audited.

Public results use tasks from Together AI's pinned evaluation archive. Counts, local predictions, and a summary reproduction script are available in the website source. External Typesafe jev-1.13 scores higher on all three sets at the default decision rule. Comparisons use different inference wrappers and do not establish general superiority. Threshold-tuned results are exploratory and are not the default results above.

Physical experiments and next research

In saved Gymnasium CartPole-v1 runs, v0.1 reached the 500-step episode limit in 10/10 episodes (mean return 500). In LunarLander-v3, its mean return was -433.8 over 10 episodes, with 0/10 reaching the score ≥200 criterion.

A separate, unreleased v0.2 DAgger experiment improved mean LunarLander return to -126.1 on the same 10 seeds, but still reached that criterion 0/10 times. Its company subset accuracy fell to 149/240 (62.08%). Those results do not describe a universal improvement or the weights in this repository. The website replays recorded simulator states, not live inference.

Current training uses labelled actions and imitation learning. Reward-based reinforcement learning, broader company tasks, threshold validation, and connecting specialist models are research directions, not completed capabilities. No real-world robot performance is claimed.

Limits and provenance

C1 can make confident mistakes, depends on the supplied options, and may misclassify legitimate messages containing words such as “spam.” Use it as an experimental decision component with human review, not an autonomous authority for financial, legal, employment, or safety-critical decisions. Performance on this small, task-specific evaluation does not establish production reliability.

C1 is an independent Qwen fine-tune by Connect AI. The option-scoring decision format is inspired by the public Together AI / Typesafe tev1 work. It is not Typesafe's jev, and is not affiliated with or endorsed by Qwen, Typesafe, or Together AI.

Adapter SHA-256: 618fe190d8408d60e7f76e3fee44d3931f9ae2326f82ebb6090c4b8c1a7e3176.

Apache-2.0; see LICENSE and NOTICE. Base weights retain their original attribution and license. Only adapter weights, inference code, and release documentation are included here.