On August 20, Harvey — the San Francisco legal AI company backed by OpenAI, Sequoia Capital and Andreessen Horowitz — published a research note announcing its first post-trained model. The notable part was not that Harvey had trained a model. It was what it started from. "Harvey Tenet is a Kimi K3 base that we post-trained together with Fireworks research for long-horizon legal work," the company wrote. Kimi K3 is the 2.8-trillion-parameter open-weight model released by Moonshot AI, a Beijing lab, three and a half weeks earlier. A company whose customers are white-shoe law firms and corporate legal departments had quietly built its flagship research model on Chinese weights — and said so out loud.
Moonshot launched K3 on July 16 and released the weights on July 27 under a modified MIT license permissive enough for commercial use. The architecture is aggressively sparse: a mixture-of-experts stack drawing from 896 experts with 16 activated per token, a 1-million-token context window, and a new attention mechanism called Kimi Delta Attention. Moonshot's launch blog claimed the combination yields "an approximate 2.5x improvement in overall scaling efficiency compared to Kimi K2." Hosted API pricing runs $3.00 per million cache-miss input tokens, $0.30 cache-hit, and $15.00 per million output tokens, flat across the full context.
One detail circulating in early coverage deserves correcting: K3 is not a mid-table model. Independent trackers placed it #2 overall on the Vals AI index and #3 on Artificial Analysis's Intelligence Index at launch, beaten only by Anthropic's Claude Fable 5 and OpenAI's GPT-5.6 Sol, and it took the top slot in Frontend Code Arena. On Artificial Analysis's cost-per-completed-task measure it comes in at $0.94, against $1.04 for GPT-5.6 Sol and $1.80 for Claude Opus 4.8. "Kimi K3 is a watershed moment because frontier open-weight models are now real," wrote Nathan Lambert of Interconnects, who had visited the Moonshot team in China. "Moonshot AI is going toe to toe with Anthropic and OpenAI with far, far fewer resources."
What Harvey Actually Built
Tenet is a research preview, not a shipped product — a distinction Harvey is careful about, writing that it is still working to "bring our work from research to production." The training run was substantial: a rank-64 LoRA across the full K3 network, adapting roughly 500,000 expert tensors, trained on approximately 150 NVIDIA B300 GPUs over two months. Harvey and Fireworks used asynchronous reinforcement learning across about 1,750 simulated legal task environments, each structured as a client matter with an expert rubric averaging 50 binary pass/fail criteria. A single rollout could span more than 1,000 turns.
The results are narrower than the headline numbers that circulated afterward. Harvey's own post says Tenet "successfully completes almost twice as many held out tasks on LAB and 20% more on LAB contracts than base Kimi K3, increasing all-pass rate by 9 and 2 percentage points, respectively." It places first on LAB: Contracts and second on LAB overall. Gains transferred to benchmarks Tenet had not trained on: on Mercor's APEX Agents corporate-lawyer subset, running K3 in Harvey's harness lifted it from a reported 58.8% to 67.5%.
Read the full post and a second finding emerges that the coverage largely missed: Tenet is not Harvey's only Chinese-weights model. For M&A due diligence — where one task can require traversing 80 million tokens of documents, and no baseline passed more than 43.8% of rubric criteria — Harvey built a recursive-language-model harness around Zhipu's GLM-5.2, lifting the pass rate to 46.1% and then 60.1% after self-distillation. Its Review Table extraction model is also a post-trained GLM-5.2. Its firm-knowledge memory model is a post-trained Alibaba Qwen3.8-27B. Harvey's research stack is Chinese open weights nearly top to bottom.
Why It Matters
The procurement question answers itself once you look at where the weights run. Harvey post-trained and serves these models through Fireworks, Baseten and its own infrastructure — American compute. Client documents never touch a Moonshot-operated endpoint. Harvey states flatly: "We did not use any customer data in any of our post-training efforts." That is the structural advantage of open weights for regulated buyers, and compliance reviewers have noted the flip side: Moonshot does not offer SOC 2 reports or a signed HIPAA business associate agreement for its hosted service, so the API and the weights are entirely different risk propositions.
The economics explain the rest. Harvey optimized Tenet for token efficiency through reward shaping, converting variable inference bills into a largely fixed cost. On its firm-knowledge work, memory cut tokens in completed trajectories by 58% and cost per query by roughly 90%. White House AI adviser David Sacks framed the move approvingly on X: "Harvey is a great example of how American companies are building world-class specialized models: they took an open-source base (Kimi K3), post-trained it on legal data, and delivered state-of-the-art performance on legal benchmarks at a fraction of the cost of frontier models."
That endorsement sits awkwardly beside the policy direction. Following K3's release, the administration reportedly revived a push to restrict Chinese AI models on cybersecurity grounds. Axios has reported that Commerce considered adding Chinese AI labs to the Entity List and that the White House weighed an executive order conditioning US hosting of Chinese models on security guarantees and breach liability. Weights already downloaded worldwide make enforcement near-impossible — but a rule would create retroactive exposure for firms that built on them.
What to Watch
Three things. Whether Tenet graduates from research preview into the production Harvey product, and whether law firms sign off when they learn the base is Chinese. Whether Harvey delivers on co-founder Gabe Pereyra's stated goal of letting firms post-train their own models on their own precedents — the shift from renting intelligence to owning it. And whether Washington acts. Alibaba has announced an open-weight Qwen 3.8 at 2.4 trillion parameters, and DeepSeek V4 is expected out of preview. If the pattern holds, the question for Western enterprise buyers stops being whether to run Chinese open weights and becomes which ones, on whose hardware, under what audit.
“Harvey Tenet is a Kimi K3 base that we post-trained together with Fireworks research for long-horizon legal work.”— Harvey research team, Harvey research post, August 20, 2026