PandaAI bets LLM agents democratize quant research; 100k users, three rounds

PandaAI builds AI-native trading research infrastructure for Chinese retail and professional quants: Qube turns a spoken idea into inspectable factor and backtest steps, EVO is an AI-native quant workbench, and an upcoming OS orchestrates multiple agents; it also runs factor competitions and open-sources QuantSkills.

LLMs will erode quant funds' research moats, so agentic research tools — factor mining, backtesting, A2A orchestration — can become the new quant infrastructure.

Li Yuqi, born in 1999, studied financial engineering at Columbia, became the youngest champion of a live futures paper-trading contest, and rose to partner at a quant private fund managing more than ¥1 billion. In 2024 he concluded that the industry's edge was largely a 'cognitive narrative' built on proprietary data, low-latency channels and large research teams — and that large language models would erode that narrative. Watching Cursor multiply programming speed without making users better researchers convinced him the bottleneck was judgment, not tooling, so he left to found PandaAI around 2024 and build AI that takes research, not returns, as its product.

PandaAI designed a three-tier product ladder: Qube for entry users (one sentence becomes data fetching, factor construction and backtests that users can inspect and reuse), EVO, which peers call the 'Claude Code of quant', as an AI-native workbench for professionals, and OS, an upcoming Agent-to-Agent environment where multiple agents iterate factors and portfolios under human-set risk rules. Underneath sits a four-layer stack: market and macro data, a self-developed CQ2 model plus DeepSeek and Doubao integrations, an ADE/A2A agent engineering layer, and live trading connections — futures desks were fully connected and a live futures client shipped in June 2026. The same month PandaAI open-sourced its first QuantSkills library and agent clusters and published research papers (CQ2, A2A Self-Evolution, AlphaSchema). By August 2026 the 70-plus-person company had 100,000+ users, had run three factor competitions drawing more than 30,000 participants, and closed seed, angel and angel+ rounds totaling tens of millions of RMB, with angel rounds led by L2F光源创业者基金; it counts brokerages, futures firms and funds as institutional partners.

Quant funds' moats — proprietary data, low-latency channels, hundred-person teams — are exactly the assets LLMs commoditize fastest.

Better coding tools did not make better researchers, which located the real gap in judgment and research process rather than execution.

An AI-native, human-in-the-loop stack avoids the scam category of autonomous 'AI stock-picking' and fits what regulators and users will accept.

Open-sourcing QuantSkills turns the obvious objection (factors leak) into a feature: agents are personalized, so the moat moves to the orchestration harness.

Running factor competitions and publishing papers converts skeptical retail users into contributors, recruitables and the core of a new methodology community.

In a zero-sum market, sell capability, not returns: PandaAI open-sources its factors and bets the moat is the agent harness — anyone can copy a factor, not an orchestration stack.

As of 2026-09-02 PandaAI is live: three rounds totaling tens of millions of RMB closed in August 2026 (angels led by L2F光源创业者基金), the futures live client shipped in June 2026, QuantSkills was open-sourced the same month, with 100,000+ users, 70+ staff and institutional partners. Li places AI trading at early L3 of five stages, heading for L4 where agents research and humans keep risk decisions; the OS layer was still upcoming. Open questions: trust in a category tied to scams, agent quality versus professional teams, and how regulators treat autonomous trading tools as they scale.