Harvey Chooses Kimi K3: A $11B Legal AI Unicorn Trains Its First Model on a Chinese Open-Weight Base

A Silicon Valley unicorn whose seed round was led by the OpenAI Startup Fund has, for the first time, trained its own model — and it did not use GPT, Claude, or Gemini. Harvey downloaded Moonshot AI's open-weight Kimi K3 and fine-tuned a legal-specialist model called Tenet (announced August 20). With its valuation rising from $11B toward a reported $15.5B, Harvey serves 50 AmLaw 100 firms plus HSBC, PwC, and KKR, consuming 12–13 trillion tokens per month. The move signals that Chinese open-weight models have shifted from being Silicon Valley's competitors to its infrastructure.

Key points

• Harvey released Tenet, its first in-house model, post-trained on Kimi K3 with Fireworks AI (async RL infrastructure) and data partners Mercor and Snorkel. No customer data was used in training.
• Why not closed APIs: per-token costs of 1/4 to 1/10 of frontier closed models; RL reward shaping penalizes unnecessary long reasoning; strategic autonomy for data sovereignty, on-premise deployment, and compliance audits required by law firms.
• Architecture fit: Kimi K3 is a 2.8T-parameter MoE model with native 1M-token context — suited to M&A due-diligence workloads reaching up to 80M tokens per task.

Training method

• Asynchronous RL with group-sequence policy optimization (GSPO), rank-64 LoRA over the full K3 network; 8 task groups × 8 rollouts per optimization step, across ~1,750 environments (>10,000 rollouts per epoch), LLM-as-a-judge using Kimi 2.6.
• 200+ practicing attorneys fabricated realistic scenarios (e.g., a three-week acquisition with a four-person diligence team); each task included a partner-level instruction, a client matter, and expert scorecards of ~50 atomic pass/fail criteria; single rollouts could exceed 1,000 turns.
• Compute: ~150 NVIDIA B300 GPUs for ~two months.

Benchmark results (self-reported, no third-party audit)

| Benchmark | Tenet vs Kimi K3 base | Type | |---|---|---| | LAB (Harvey's open legal agent benchmark, 1200+ tasks) | ~2× task completion, all-pass +9pp | Internal | | LAB Contracts | +20% completion, all-pass +2pp, SOTA | Internal | | Mercor APEX Agents (Corporate Law) | Significant improvement | Cross-benchmark transfer | | Crosby Redline Bench | Significant improvement | Cross-benchmark transfer | | Scale PRBench hard | 36.0% → 36.8% (not statistically significant, per Harvey) | Cross-benchmark transfer | | LegalBench / CUAD / MAUD | No regression | Knowledge benchmarks |

Specialist models on top of Tenet

• M&A Diligence: Recursive Language Model architecture (root agent holds a database in a REPL, dispatching subtasks); a GLM-5.2 orchestrator alone scored 46.1% pass rate, 60.1% after fine-tuning, vs a 43.8% ceiling for baseline models.
• Review Table: post-trained on GLM-5.2; answer quality +3.6, citation quality +12.1, per-cell cost down to ~1/10.
• Firm Knowledge: post-trained on Qwen3.8-27B; pass rate +15%+, tokens per trajectory −58%, per-query cost −~90%.

The broader pattern: Kimi bases powering US AI products

• K2.5 → Cursor Composer 2 (and Composer 2.5)
• K2.6 → Cosine Lumen Outpost (legacy-code maintenance, UK)
• K2.7 → Devin SWE-1.7 (Cognition, July 2026; FrontierCode score 45, matching Opus 4.8 at 1/4 the cost)
• K2.5 synthetic data → Thinking Machines Inkling (Mira Murati's $12B+ startup)
• K3 → Harvey Tenet

Supporting platform data: Chinese open-weight models' share of OpenRouter token traffic rose from ~1.2% to ~30% in a year; Vercel AI Gateway reported K3 daily tokens up ~3× and Kimi-family tokens up 4×+ in July; GitHub Copilot added K3 on August 6 (20M+ users); Arena's coding leaderboard placed K3 second at 1660 vs Claude Opus 5's 1699. a16z's Anjney Midha estimates ~80% of top US AI companies use Chinese open models.

Harvey's economics

Harvey's ARR grew from $100M (Aug 2025) to $190M (end of 2025) to ~$300M (May 2026), with 1,300+ clients, 200,000+ lawyers, 167% net revenue retention, and 25,000+ custom agents running in client organizations. Monthly token spend is described as in the billions of dollars — making API costs an existential margin problem. The trigger: OpenAI hired Ironclad's founder to lead a legal-industry unit, and Anthropic launched lawyer-facing tools — Harvey's biggest suppliers became its competitors.

Silicon Valley's open-source shift

OpenAI released open-weight GPT-OSS models; AWS added GPT-OSS and Nemotron to Bedrock; NVIDIA launched the Nemotron Coalition (March 2026) and struck a ~$6B deal with Poolside (August 2026) including $1B investment and 100+ engineers. Meta, Microsoft, Palantir, Google, and OpenAI signed a July industry letter supporting open-weight AI — Anthropic notably did not.

Caveats

• Tenet is a research preview — no public weights, model card, or API endpoint yet.
• Results are largely self-reported; training tasks and scorecards are proprietary and hard to reproduce.
• Using open-weight Kimi K3 involves no commercial relationship with Moonshot AI — monetization, licensing, and geopolitical compliance questions remain unanswered.
• Fast base-model iteration (e.g., Alibaba's Qwen releases) means Harvey faces the same vendor-lock question it tried to escape.

Conclusion

Harvey's story is a landmark "build vs buy" case: unable to afford training from scratch and unwilling to stay locked into closed APIs, it took a Chinese open-weight base and fine-tuned it with its own legal RL data. When an OpenAI-backed unicorn bets on Kimi K3, the industry's national boundaries blur while cost curves and data sovereignty are redrawn. Harvey will not be the last.

Sources: Harvey official announcement (2026-08-20); 21st Century Business Herald (2026-08-24); World Programming; Cognition official X (2026-07-08); OpenRouter / Vercel AI Gateway traffic data; GitHub Copilot announcement (2026-08-06); Arena leaderboard (2026-08-19); NVIDIA Nemotron Coalition and Poolside announcements.