{"slug": "harvey-built-its-own-legal-ai-model-instead-of-renting-one-from-openai", "title": "Harvey Built Its Own Legal AI Model Instead of Renting One From OpenAI", "summary": "Harvey, the $11 billion legal AI startup, unveiled Tenet, its first proprietary in-house model, as part of the Harvey II relaunch on August 18. Tenet, built on a customized version of Moonshot AI's open-weight Kimi K3 model and post-trained with Fireworks AI, improved Harvey's internal LAB benchmark all-pass rate by 82% and LAB Contracts by 22%, while running at under a quarter of the token cost of frontier models from OpenAI and Anthropic. The move reflects a broader industry shift toward owning specialized models to reduce reliance on external AI providers.", "body_md": "*Harvey, the $11 billion legal AI startup, just stopped renting all of its intelligence from OpenAI and Anthropic. It built its own model, called Tenet, on top of an open-weight Chinese model, and it's letting law firms fine-tune it with their own case files.*\n\nHarvey unveiled Tenet this week as the centerpiece of a broader relaunch called Harvey II. Tenet is Harvey's first proprietary, in-house model. It starts from a customized version of Kimi K3, the open-weight model from China's Moonshot AI. Harvey then post-trained that base with Fireworks AI for what the company calls long-horizon legal reasoning: the multi-hour, multi-document slog of reviewing contracts or building a litigation record, not a quick chat answer.\n\n## The training bet\n\nThe training data is the interesting part. Harvey hired attorneys through Mercor and Snorkel, two platforms that supply specialized human labor for AI training. It put them to work on mock disputes and synthetic case files. The goal: mimic real legal work. Combined with publicly available legal data, that human-generated corpus is what separates Tenet from a generic foundation model.\n\nAccording to Harvey's own technical writeup, the post-training lifted Tenet's all-pass rate by 82% on its internal LAB benchmark and 22% on LAB Contracts, compared with the base Kimi K3 model. That was enough. It put Tenet in first place on LAB Contracts and second on LAB overall. Harvey also says Tenet runs at under a quarter of the token cost of the frontier models it used to depend on exclusively.\n\nThat cost math is the actual business case here. Legal AI products often run for hours on a single matter, chewing through discovery documents, prior filings, and internal memos. Every one of those tokens used to flow through an API call to OpenAI or Anthropic, priced at frontier rates. Owning a fine-tuned, cheaper model for the bulk of that work changes Harvey's margin structure on every hour of legal work its software touches. It presumably still calls frontier labs for the hardest cases.\n\n[Moonshot AI closes round at $31.5 billion and is already chasing $50 billion](https://startupfortune.com/moonshot-ai-closes-round-at-315-billion-and-is-already-chasing-50-billion/)\n\nMoonshot AI, the Beijing-based lab behind Kimi K3, closed a funding round at a $31.5 billion valuation in late July and is already in talks to raise again at up to $50 billion ahead of a Hong Kong IPO. Its 2.8-trillion-parameter open model shipped inside Cursor on launch day, and its pricing undercuts OpenAI and Anthropic by up to 9x, accelerating... - [Chinese AI startup valuation milestone](https://startupfortune.com/moonshot-ai-closes-round-at-315-billion-and-is-already-chasing-50-billion/) - [Moonshot AI funding round 31.5 billion](https://startupfortune.com/moonshot-ai-closes-round-at-315-billion-and-is-already-chasing-50-billion/)\n\nBloomberg Law reported this week that legal tech firms broadly are shifting away from pure reliance on OpenAI and Anthropic. Harvey's move is the clearest example yet. It fits a pattern playing out across vertical AI: once a startup has enough proprietary data and enough transaction volume, running a frontier model's API for every request stops making financial sense. Harvey isn't abandoning the big labs. It's routing the routine, expensive-to-serve work to something it owns and controls.\n\n## A model each firm can shape\n\nHarvey is also letting individual law firms fine-tune Tenet on their own institutional knowledge. A firm's specific drafting conventions, past matters, and internal precedent can now shape the model's output, rather than relying on Harvey's general training alone. That's a meaningfully different pitch than \"we plugged into GPT-5.\" It's an offer to build a firm-specific instrument.\n\nTenet arrives bundled into Harvey II, the platform overhaul the company launched on August 18. The headline feature there is Memory, which learns an individual lawyer's drafting style, tone, structure, and citation habits, and carries those preferences across Harvey, Microsoft Word, and Outlook. The rollout is staged. Personal Memory, focused on one lawyer's habits, is in early access now. A second phase will extend memory across a shared matter or client vault. A later phase will let entire firms bring shared conventions into the system, while keeping boundaries between clients and teams intact. Harvey says Memory data won't be used to train its global models, and users can review, edit, or turn the feature off.\n\n## The money behind the model\n\nNone of this is happening in a vacuum. Harvey closed a $200 million round in March at an $11 billion valuation, led by existing backers GIC and Sequoia. SiliconANGLE reported this month that the company is now in talks to raise as much as $500 million at a $15.5 billion valuation. Its annualized revenue has reportedly jumped from roughly $190 million in January to more than $350 million by August, according to PYMNTS. Rival Legora, meanwhile, has been raising at a reported $5.5 billion valuation. The two are increasingly racing on the same turf: who can make AI-assisted legal work fast enough, and cheap enough, to become the default at big firms.\n\nFrankly, the choice of Kimi K3 as the base is the detail worth sitting with. A well-funded American legal AI company, backed by Sequoia and GIC, is building its core infrastructure on an open-weight model out of China, rather than paying OpenAI or Anthropic by the token. That's not a story about ideology. It's a story about what happens once open-weight models get good enough that a company with real data and real volume would rather own its stack than rent someone else's.\n\n**Also read:** [Fortinet Buys Virtue AI to Hunt Vulnerabilities in AI Agents Before Hackers Do](https://startupfortune.com/fortinet-buys-virtue-ai-to-hunt-vulnerabilities-in-ai-agents-before-hackers-do/) • [NSA and CISA Warn Hackers Are Using AI to Scan Siemens Plant Controllers](https://startupfortune.com/nsa-and-cisa-warn-hackers-are-using-ai-to-scan-siemens-plant-controllers/) • [Micro1 Rockets From $7 Million to $300 Million in Revenue in a Single Year](https://startupfortune.com/micro1-rockets-from-7-million-to-300-million-in-revenue-in-a-single-year/)", "url": "https://wpnews.pro/news/harvey-built-its-own-legal-ai-model-instead-of-renting-one-from-openai", "canonical_source": "https://startupfortune.com/harvey-built-its-own-legal-ai-model-instead-of-renting-one-from-openai/", "published_at": "2026-08-21 02:36:39+00:00", "updated_at": "2026-08-21 02:43:56.114305+00:00", "lang": "en", "topics": ["artificial-intelligence", "large-language-models", "ai-products", "ai-startups"], "entities": ["Harvey", "Tenet", "OpenAI", "Anthropic", "Moonshot AI", "Kimi K3", "Fireworks AI", "Mercor"], "alternates": {"html": "https://wpnews.pro/news/harvey-built-its-own-legal-ai-model-instead-of-renting-one-from-openai", "markdown": "https://wpnews.pro/news/harvey-built-its-own-legal-ai-model-instead-of-renting-one-from-openai.md", "text": "https://wpnews.pro/news/harvey-built-its-own-legal-ai-model-instead-of-renting-one-from-openai.txt", "jsonld": "https://wpnews.pro/news/harvey-built-its-own-legal-ai-model-instead-of-renting-one-from-openai.jsonld"}}