News / Model Watch · Ecosystem Adoption
OpenAI-Backed Harvey Builds Its First Model on Kimi K3 Open Weights
The strongest validation yet for Kimi K3 as a production base: Harvey — the San Francisco legal-tech startup backed by OpenAI, Sequoia, and a16z — announced Harvey Tenet on August 20, its first post-trained model. The base? Moonshot's 2.8T open weights. The result: roughly 2× the completed tasks of base K3 on Harvey's Legal Agent Benchmark, at what the company calls open-source cost.
Harvey Tenet is a Kimi K3 base, post-trained with Fireworks through asynchronous RL for long-horizon legal work, per Harvey's official announcement (Aug 20, 2026). It's a research preview: ~2× held-out task completion vs base K3 on Harvey's LAB, ~20% more on related agentic evals, frontier-level legal benchmarks — but no published weights, model card, or API. The base is open; Tenet itself is Harvey's checkpoint and will ship inside Harvey's products.
What shipped, per Harvey's announcement
- The model. Kimi K3 base + asynchronous reinforcement learning over synthetic data, publicly available legal data, and human expert data — explicitly no customer data. Trained together with Fireworks AI's research team.
- The results. Nearly 2× held-out tasks on Harvey's Legal Agent Benchmark vs the K3 base; ~20% gains on related agentic evaluations; harness improvements on top for training and task execution.
- The positioning. "Frontier-level on prominent legal benchmarks… at an open-source cost," making continuous agent runs across every matter practical — plus the groundwork for law firms to build specialized models on their own work.
- What didn't ship. No weights, no model card, no API endpoint — research preview only. As MarkTechPost put it: "what ships today is the recipe, not the artifact."
Why it matters: the six-month backstory and the sovereignty angle
- Open weights finally caught up. Harvey tried post-training before and stopped: OpenAI's base models improved so fast that customizations were leapfrogged, and earlier open models were too far behind the frontier (per community discussion of the company's history). K3 changed the calculus.
- A Western, OpenAI-backed firm choosing Chinese open weights. SCMP frames it as part of a growing shift by Western tech firms toward Chinese open-weight systems amid soaring development costs.
- The "own your intelligence" pitch. Harvey's stated goal includes letting law firms build their own specialized models — two firms using Harvey end up with different models shaped by different work. That's an AI-sovereignty argument built on open weights, not proprietary APIs.
Primary sources
- Harvey 官方博客 — Update on Harvey's Post-Training Effort（2026-08-20，Tenet = Kimi K3 base + Fireworks）
- Law.com LegalTechNews — Harvey Introduces Tenet, Its First Post-Trained AI Model for Legal（2026-08-20）
- SCMP — OpenAI-backed legal tech firm pivots to Chinese Kimi K3 open-weight model
- MarkTechPost — Harvey Tenet: A Kimi K3 Base Post-Trained with Fireworks（2026-08-23）
- Artificial Lawyer — Harvey Tenet 分析（2026-08-21）
FAQ (2026)
What is Harvey Tenet?
Harvey's first post-trained model (research preview, Aug 20, 2026): Kimi K3 open-weight base + Fireworks async-RL post-training for long-horizon legal work.
vs base Kimi K3?
~2× held-out tasks on Harvey's Legal Agent Benchmark, ~20% more on related agentic evals — frontier-level legal benchmarks at open-source cost, per Harvey.
Is Tenet open-weights?
No — the base is; Tenet is Harvey's checkpoint. No weights, model card, or API at announcement; it ships inside Harvey's products.
Why Kimi K3?
Six months of open-weight research: earlier bases either leapfrogged too fast or lagged the frontier. K3's 2.8T/1M-context open weights finally held still long enough to build on.
Trained on what?
Async RL over synthetic + public legal + human expert data; no customer data. Built with Fireworks AI's research team.
Why does it matter?
An OpenAI-backed Western firm's first in-house model on Chinese open weights — the clearest adoption signal yet for K3 as a domain post-training base.