OpenAI-backed Harvey builds first model on Chinese base Kimi K3
新智元 · wechat · 2026-08-25
Harvey, a $11B legal AI company backed by OpenAI, released its first self-trained model, Harvey Tenet, built on China's open-source Kimi K3 base instead of US closed-source models.
Why Kimi K3?
- Cost & Control: Harvey processes 13T tokens/month. Relying on APIs was costly and risky as vendors (OpenAI/Anthropic) enter the legal space.
- Performance & Editability: As the largest open-weights model (2.8T params MoE), K3 allows deep fine-tuning (LoRA). Tenet costs <1/4 of top models but outperforms on legal benchmarks.
- Context: 1M token window handles massive legal documents.
Training Method
Harvey trained Tenet in 2 months using 150 B300 GPUs via "asynchronous reinforcement learning". Real lawyers fabricated cases and strictly graded outputs to train the model.
Industry Trend
Following Cursor and Perplexity, Harvey's choice signals that open-weights models are becoming the preferred base for vertical applications seeking cost control and differentiation.
Related event: Harvey's First In-House Legal Model Tenet Built on China's Kimi K3(2 posts)→
More from Companies & People
- Cohere expands RL Environments team training agents in simulated enterprise setups — cohere · 2026-10-16
- Vector Institute partners with Town of Caledon in its first municipal AI adoption deal — VectorInst · 2026-10-16
- Follow-up: even small X accounts got into Google's Argon EAP — lyraxana · 2026-10-16
- Zuck posts AI agent bought him a watermelon; Threads users roast him mercilessly — venturetwins · 2026-10-16
- PyTorch Conference NA 2026: Oct 20-21 in San Jose, 3,500+ attendees, 200+ talks — PyTorch · 2026-10-16
- NYC Comptroller urges Anthropic to pause IPO until AI safety and governance concerns are addressed — ShakeelHashim · 2026-10-16