What is Qlib?
Qlib is Microsoft Research's platform for AI-driven quantitative investment research. It covers the chain from a point-in-time financial data layer through model training and backtesting to portfolio optimisation, order execution and online model serving, and supports supervised learning, market-dynamics modelling and reinforcement learning.
The examples/benchmarks folder holds 25 model implementations, from LightGBM, XGBoost and CatBoost to LSTM, GRU, Transformer, TabNet, TRA, HIST and DoubleEnsemble, each runnable from a YAML config with the qrun command. LLM-driven factor and model search lives in a sister project, RD-Agent, which the README promotes as Qlib's automated R&D layer.
The repo has about 50,300 stars and an MIT licence. The latest GitHub release, v0.9.8, came out in October 2026 as a security-hardening release; PyPI still serves pyqlib 0.9.7 from August 2025.
Is Qlib worth it?
Qlib earns its place as the shared vocabulary of open-source ML quant research. Its Alpha158 factor set, dataset handlers and backtest are what other projects build on; vn.py's alpha module borrows Alpha158 outright and credits Qlib for its design. Having 25 benchmark models behind one config format means a new idea can be compared against LightGBM, an LSTM or a Transformer baseline in an afternoon rather than a week.
The weak point is data. The official dataset download is disabled, temporarily according to the README, under a stricter data security policy, and users are pointed to a community-maintained mirror instead. The bundled collectors scrape Yahoo Finance, which the maintainers warn might not be perfect. Serious users will convert their own vendor data with the CSV tools; anyone hoping for a clean, ready dataset will be disappointed.
Releases have slowed to about one a year: v0.9.6 in December 2024, v0.9.7 in August 2025 and v0.9.8 in October 2026. The latest is about security, not features. It replaces evaluation of feature expressions with a restricted parser, and loading saved recorder artefacts or local Python modules from configs now needs an explicit trusted flag. That breaks reloading of older models, but it closes a real hole: before it, a shared experiment folder or config file was a path to running someone else's code. About 516 issues and pull requests are open.
The LLM layer sits outside. RD-Agent, the MIT-licensed multi-agent system that mines factors and tunes models against Qlib, has about 15,400 stars of its own and needs an LLM API key and a compute budget. Qlib itself calls no model provider. As a research bench for a quant team or a graduate student it remains the strongest free option; as a live trading stack it needs a broker layer, risk controls and your own data on top.
Where it earns its place
A two-person quant team at a family office
Test whether a new factor adds anything over standard baselines
Convert their vendor data with the CSV tools, add the factor to Alpha158, and run LightGBM and Transformer configs with qrun; the backtest report compares excess return with and without it. Free apart from compute.
A finance master's student
Reproduce a published deep-learning stock model for a thesis
Start from the matching folder in examples/benchmarks, which fixes the model, dataset and backtest settings, then change one variable at a time. Tree models run on a laptop; the neural models want a GPU.
An ML engineer at a systematic fund
Let an LLM agent propose and test factors overnight
Point RD-Agent at a Qlib workspace with an LLM API key; it writes factor code, runs Qlib backtests and keeps what improves the score. API spend grows with iterations, and every promoted factor still needs a human review.
Before you commit
- Setup
- pip install pyqlib on Python 3.8 to 3.12 (conda recommended), or build from source to get v0.9.8; then download the community data mirror or convert your own CSV data before qrun does anything useful.
- Integrations
- Python API and the qrun CLI; LightGBM, XGBoost, CatBoost and PyTorch models; Yahoo Finance collectors; RD-Agent for LLM-driven factor and model search.
- Data and privacy
- Runs on your own machine and calls no model provider. Since v0.9.8, saved artefacts and local Python modules load only with an explicit trusted flag, which should be granted only to storage you control.
- Cost at scale
- No paid edition. Costs are your market data licence, GPU time for the neural models and, if you add RD-Agent, LLM API spend.
- Limits
- About 516 open issues and pull requests; official dataset paused; Yahoo-sourced data may be imperfect; the high-frequency trading example is only partly released; no live broker connections.
What’s good
- 25 benchmark models runnable from one YAML config format
- Full research chain: data layer, training, backtest, portfolio and execution
- Reinforcement learning framework alongside supervised models
- October 2026 release closes code-execution paths in configs and artefacts
- MIT licence
What’s not
- Official dataset download disabled; default collectors scrape Yahoo Finance
- About one release a year, and PyPI lags GitHub
- v0.9.8 trust flags break reloading of older artefacts
- No broker connections; live trading needs another layer
Best for
Quant researchers and students who want a tested ML pipeline and baselines for stock prediction and will supply their own clean data.
Qlib alternatives
Qlib: common questions
- Is Qlib worth it?
- Still the reference open-source framework for machine-learning stock research, provided you bring your own data and treat it as a research bench rather than a trading system. We rate it 4.1 out of 5.
- How much does Qlib cost?
- Free and open source under MIT. No paid edition; you pay for data, compute and, with RD-Agent, LLM API usage.
- What are the downsides of Qlib?
- Official dataset download disabled; default collectors scrape Yahoo Finance; about one release a year, and PyPI lags GitHub; v0.9.8 trust flags break reloading of older artefacts; no broker connections; live trading needs another layer.
- Who is Qlib best for?
- Quant researchers and students who want a tested ML pipeline and baselines for stock prediction and will supply their own clean data.
- How much does Qlib cost for a team or in production?
- No paid edition. Costs are your market data licence, GPU time for the neural models and, if you add RD-Agent, LLM API spend.
- Is Qlib safe for company data?
- Runs on your own machine and calls no model provider. Since v0.9.8, saved artefacts and local Python modules load only with an explicit trusted flag, which should be granted only to storage you control.
- What are the best Qlib alternatives?
- The closest alternatives we have reviewed are Kronos, VeighNa (vn.py) and Vibe-Trading.
Answers checked Nov 16, 2026.