Jingyao Gong
Jingyao Gong is an individual student, researcher, and open-source AI developer rather than a company. He publishes reproducible, readable implementations of small language, vision-language, and omni models so that learners can understand and train models from scratch on comparatively modest hardware.
At a Glance
- Students and AI/ML learners
- Open-source developers and researchers
- Small-model and efficient-AI researchers
- Developers experimenting with local language, vision-language, and speech models
AI Tools by Jingyao Gong
(1)
MiniMind
Train LLM From Scratch PyTorch
Discussions
No discussions yet
Be the first to start a discussion about Jingyao Gong
Latest News
MiniMind-O first release: minimind-3o and minimind-3o-moe, with text/speech/image input and streaming speech output
MiniMind-V 3v and 3v-MoE update with new checkpoints, SigLIP vision encoder, projector and dataset pipeline changes
MiniMind-3 and MiniMind-3-MoE release with Qwen3-aligned architecture, new tokenizer, Agentic RL and tool-calling support
MiniMind update adding native PPO/GRPO/SPO-style RLAIF work, checkpoint resume, YaRN, adaptive thinking and tool-calling support
Products & Services
Apache-2.0 open-source small language-model project and tutorial. It provides native-PyTorch code for tokenizer training, pretraining, SFT, LoRA, DPO, PPO/GRPO/CISPO, tool use, Agentic RL, adaptive thinking, distillation, evaluation, and inference. Current mainline models include a 64M dense model and a 198M/A64M MoE model.
Vision-language extension supporting image understanding and conversational responses. The series includes 26M/104M MiniMind2-V models and 65M/200M-A65M MiniMind-3V models, with training and evaluation code.
Small omni model with text, speech, and image inputs and text plus streaming speech output. It uses a Thinker-Talker design, Mimi audio codebooks, multi-token audio prediction, voice cloning, real-time barge-in, and phone-mode/WebUI demos; dense and MoE releases are approximately 0.1B and 0.3B-A0.1B.
A diffusion-language-model extension listed by the MiniMind project as part of its broader model family.
Market Position
MiniMind is positioned as a transparent, low-cost, educational alternative to heavyweight model stacks and black-box APIs: it prioritizes source-level readability, full reproducibility, and training from scratch on one consumer GPU. MiniMind-O's report situates its omni work alongside Mini-Omni/Mini-Omni2, Moshi, GLM-4-Voice, Qwen3-Omni, LLaMA-omni, and related speech/omni systems, while emphasizing a much smaller and more inspectable implementation.
Leadership
Founders
Jingyao Gong
Student and open-source AI developer. His personal site identifies him as a student in Jiangsu, China; an author bio says he is pursuing a master's degree in Computer Technology at Jiangsu University's School of Computer Science and Communication Engineering, with interests in image processing, computer vision, and deep learning. His GitHub profile also lists Nanjing University, so the public profiles are not fully consistent about institutional affiliation.
Founding Story
Gong started MiniMind to lower the barrier to understanding large language models: instead of only using high-level libraries or fine-tuning existing checkpoints, learners should be able to read the code and train a small model from zero. The project emphasizes a complete, native-PyTorch pipeline that can run on an ordinary personal GPU, initially advertised as training a roughly 64M-parameter chatbot in about two hours for about RMB 3 on one RTX 3090.
Business Model
Revenue Model
The projects are released as free Apache-2.0 open-source software and model/code/data releases. No commercial monetization model is stated; the RMB 3 figure is an example GPU rental cost for training, not a product price.
Target Markets
- Students and AI/ML learners
- Open-source developers and researchers
- Small-model and efficient-AI researchers
- Developers experimenting with local language, vision-language, and speech models
- Learning how LLMs, VLMs, and omni models work from source code
- Reproducing small-model training experiments on consumer GPUs
- Research and experimentation in language modeling, multimodal learning, speech interaction, RLHF/RLAIF, and distillation
- Building lightweight local chat, tool-use, vision, and speech demos
- Teaching and hands-on tutorials for model training