Open Source Project MiniMind Gains Approximately 60,000 Stars in Two Years
The GitHub project MiniMind features around 64M parameters, takes about two hours on a single NVIDIA 3090 card, and costs approximately 3 RMB for card rental to train a conversational small model from scratch. The repository was created around July 2024 and has garnered about 60,000 stars.
The author account jingyaogong is associated with Nanjing University, with submissions highly concentrated from one individual. The documentation specifies that the "two hours, three yuan" corresponds to one epoch of SFT on a single 3090 card, based on actual measurements and rental prices, not the total cost for the entire process; pre-training alone takes about 1.21 hours and costs approximately 1.57 RMB. The main model minimind-3 consists of 8 layers, with a hidden dimension of 768 and a vocabulary of 6400, aligned with Qwen3; there is also an MoE version with a total of about 198M parameters and approximately 64M active parameters. The implementation uses native PyTorch to cover pre-training, SFT, LoRA, DPO, PPO/GRPO/CISPO, distillation, tool invocation, and Agentic RL, without relying on high-level wrappers like transformers, trl, or peft. Inference can connect to llama.cpp, vLLM, and ollama, and provides services compatible with OpenAI protocols and web chat. Derivatives include MiniMind-V, MiniMind-O, and diffusion language model branches.
64M is roughly 1/2700 of the parameter count of GPT-3, capable of basic multi-turn interactions and simple tool invocation, but not suitable for cutting-edge model tasks. The project shifts the teaching focus from "being able to tune LoRA" to "being able to complete a training loop."
In terms of market mechanisms, this is an open-source teaching repository diverting traffic from cloud-based fine-tuning courses. Beneficiaries include individual developers who need to understand the loop and short-term orders from rental card platforms; the pressure falls on tutorial providers that only sell packaged fine-tuning templates. Stars represent attention, not income. Funding remains at around 1.3 RMB per hour for 3090 rentals, not entering new foundational model companies.
The growth rate of stars proves the topic's validity. The model's capability boundary is still determined by 64M, and a full-link script does not equate to a commercially viable foundation.
Source: Public Information
ABAB AI Insight
Writing a training loop that can run on a consumer card serves as a counter to the notion that "large models can only be understood through papers." The author’s nearly solo submissions indicate that such projects still rely on individuals to break down algorithms from libraries back into original operators. The alignment with Qwen3 reduces friction in vocabulary and inference ecosystems but also ties the small model to existing Chinese ecosystems. The three yuan hook sells the SFT for one hour, while the full link including pre-training and reinforcement learning would return to costs of dozens of yuan and dozens of hours, as the documentation itself narrows the scope.
Capital will not value 64M at 60,000 stars. It will value labor based on whether engineers truly understand GRPO. High-level wrapper courses continue to sell speed, while native implementation repositories sell readability, creating two parallel markets. Derivative visual and multimodal models are extensions of the same author's courses, rather than data center-level pre-training.
In contrast to nanoGPT, Karpathy's courses, and various schools writing Transformer assignments from scratch, the industry has shifted from API calls to reproduction cycles: those who can complete alignment algorithms on a single card qualify to modify algorithms.
This represents a technological substitution: black-box fine-tuning courses are being replaced by a minimally operational closed loop. The mechanism is to trim parameters to fit consumer cards while retaining the full algorithm, allowing cost narratives to spread.
ABAB News · Cognitive Laws
Loops that can run on a single card are closer to the original algorithm than merely knowing how to fine-tune.
The two hours and three yuan refer to one phase, not the entire training history.
Stars measure how understandable the topic is, not how strong the model is.