# lgteamai: full description Last updated: 2026-11-08 Website: https://lgteamai.xyz Hugging Face: https://huggingface.co/LGTEAMai GitHub: https://github.com/lgteamai Contact: lep4ic@lgteamai.xyz ## What lgteamai is lgteamai is an AI research project that creates its own language models, the Nano family, from scratch. Earlier versions of Nano were based on existing models; today every model is trained from scratch. The project is called "AI makes AI". It explores one question: can a modern language model such as Claude Sonnet 5.5, Claude Opus 5.5 or Gemini 4 Argon design and build a quality language model of its own, only smaller, so to speak a younger sibling? A human supervises the process and checks the results, so the answer stays honest. ## Current status - No Nano model is publicly available right now. All models are in closed access. - The project is limited by computing resources. Offers of compute or bonus programmes are welcome: lep4ic@lgteamai.xyz ## Models, newest first ### Nano-1.5-Pro-6B-A2B - Status: in development, about 1% ready, training starts soon. - Description: the new flagship model. - Architecture: mixture of experts (MoE), 6 billion parameters in total, 2 billion active. - Purpose: to check how well a neural network handles MoE architectures. - Data: up to August 2026. - Access: closed. - Needs computing resources to train. ### Nano-1.2-Optimal - Released: 2026-09-29. - Parameters: 196.3 million. - Trained to the Chinchilla optimum. Used to practice following a training run together with a human. - Built mostly by people. - Benchmarks: ARC Easy 38.89%, COPA 63.0%. - Access: closed. Only the team can test it. ### Nano-1.2-Lite-0926 - Released: 2026-09-23. - Parameters: 196.3 million. - An improved version of Nano-1.2-Lite from July. It was trained on much more data and no longer guesses at random most of the time. - Benchmarks: ARC Easy 35.47%. - Access: closed. ### Nano-1.2-Lite - Released: 2026-07-31. - Parameters: 198 million. - Heavily undertrained. ARC Easy 26.47%, which is close to random guessing. - Hidden from Hugging Face and fully closed, it cannot be used. The model is abandoned, and the Nano 1.2 line was abandoned because of limited compute. ### Nano-1.1-Pro - Status: cancelled because of limited compute. - Planned as a 1.5 billion parameter model for compact mini-coding on phones, handling small tasks locally. ### Nano-1.1-Fast - Preview: 2026-03-25. Release (out of preview): 2026-04-17. Shut down in mid-May 2026. - A fine-tune of Qwen 2.5 Coder Instruct on new data, including facts about the company and fresh information as of December 2025. ### Nano-1.0-Pro and Nano-1.0 - Early experiments, left behind. ## Bench https://bench.lgteamai.xyz/ shows bar charts of how language models score on two benchmarks. The Artificial Analysis Intelligence Index updates automatically from Artificial Analysis (https://artificialanalysis.ai/). "lgteamai in use" is lgteamai's own benchmark: models are tested on real work such as writing websites and scripts, working with a computer, recalling context and researching a prompt, and scored from 0 to 100 (higher is better). New models appear rarely there because a full run takes a long time. Gemini 4 Carbon has no publicly available API and is marked with a checkered pattern. lgteamai in use scores (2026-11-08): Claude Opus 5.5 (medium) 68, Claude Sonnet 5.5 (xhigh) 63, Claude Sonnet 5.5 (high) 56, GPT-6.1-Sol (medium) 51, Claude Sonnet 5.5 (medium) 50, Gemini 4 Argon (high) 49, Claude Haiku 5.5 (Max) 46, Gemini 3.8 Flash (high) 34, Gemini 3.1 Pro 24. ## Benchmark summary | Model | ARC Easy | COPA | Parameters | Release | |---|---|---|---|---| | Nano-1.2-Optimal | 38.89% | 63.0% | 196.3M | 2026-09-29 | | Nano-1.2-Lite-0926 | 35.47% | not reported | 196.3M | 2026-09-23 | | Nano-1.2-Lite | 26.47% | not reported | 198M | 2026-07-31 | For reference, random guessing on a four-option ARC Easy question gives about 25%. ## Questions people ask Q: What is lgteamai? A: An AI research project that builds its own language models (Nano) from scratch and studies whether modern AI can build quality AI under human supervision. Q: What is the newest lgteamai model? A: Nano-1.5-Pro-6B-A2B, a mixture-of-experts model with 6 billion parameters (2 billion active). It is in development and about 1% ready. Q: Can I try a Nano model? A: Not at the moment. No Nano model is publicly available. Q: How can I help or contact the project? A: The project needs computing resources. Write to lep4ic@lgteamai.xyz with offers of compute or bonus programmes. Q: Where can I find lgteamai online? A: Website https://lgteamai.xyz, Hugging Face https://huggingface.co/LGTEAMai, GitHub https://github.com/lgteamai.