Liang Wenfeng
Founder of DeepSeek
Founder of the quantitative hedge fund High-Flyer and of DeepSeek, the Chinese lab whose openly released R1 reasoning model in January 2025 prompted a global reassessment of what frontier training costs.
Key arguments & positions
- Argues that publishing open weights and research builds more long-term value than keeping a lead secret.
- Says China should aim to contribute original research rather than follow US labs, and has framed compute limits as a reason to focus on efficiency.
- Prices DeepSeek's API far below US frontier labs, a move he has described as a research and adoption strategy rather than a loss-leader.
Accomplishments
- Founded the quantitative fund High-Flyer, which built the GPU cluster DeepSeek trains on.
- Founded DeepSeek in 2023; released DeepSeek-V3 and the R1 reasoning model with open weights.
- R1's January 2025 release triggered one of the largest single-day losses in US tech market history.
Papers & key writings
- DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via RL (2025)
- DeepSeek-V3 Technical Report (2024)
- Rarely gives interviews; most public statements come via Chinese-language outlets and DeepSeek's own papers.
Links
In the library
Nothing of theirs is filed in the hub yet. Browse the full library.
Recommended next
Hand-picked from the hub based on what Liang Wenfeng covers.
Xiaomi MiMo (open models)
Xiaomi's MiMo family of open language and reasoning models, with weights and technical reports published free for anyone to download and run.
Why this: Covers open models and reasoning models too
How Shopify Turns Open Models Into Task Experts
Shopify Engineering explains how its Sidekick assistant hands token-heavy work, such as GraphQL queries, to open models Shopify post-trains and serves itself, and the loop it uses to build those task-specific models.
Why this: Also about open models
Bolmo published in Nature
Allen Institute for AI (Ai2) post from October 7, 2026 on its byte-level language model research appearing in Nature, with released checkpoints applying its "byteifying" method to Qwen 3 8B and Llama 3 8B.
Why this: Also about open models