A week ago, U.S.-based legal tech company Harvey made an announcement on X, unveiling its first proprietary model, Harvey Tenet, trained on Kimi K3. The news quickly stirred waves in the AI community.

The practice of post-training on open-source models is not new. In March, Cursor's proprietary programming model, Composer 2, was revealed to be trained on Kimi K2.5. Companies like DoorDash and Airbnb have also adopted open-source models to cut costs.

Why has a legal tech company's 'proprietary model' garnered so much attention domestically and internationally? The answer lies in the fact that Harvey might be the last AI company one would expect to choose a Chinese model.

Founded in 2022, Harvey has secured investments from OpenAI, Sequoia Capital, a16z, and others, reaching a market valuation of $11 billion. Over 75 Am Law 100 firms globally are Harvey's clients. Harvey's product prototype was built on OpenAI's models, and the two companies even collaborated to train custom models for U.S. case law.

Despite being seen as OpenAI's 'prodigal son,' Harvey chose a Chinese open-source model as the foundation for its first foundational model.

Harvey's relationship with OpenAI extends beyond the typical 'investor-investee' dynamic.

One founder, a former securities and antitrust lawyer, and the other, with experience at DeepMind and Meta, initially used GPT-3 to test legal Q&A before securing $5 million in seed funding led by the OpenAI Startup Fund.

In Harvey's early days, OpenAI served as both a capital source and a crucial technological backbone.

By 2023, when Harvey completed its $80 million Series B funding round, OpenAI and Sequoia Capital continued to participate. Harvey explicitly stated in its funding announcement that it was collaborating with OpenAI to develop domain-specific models for the professional services industry.

The division of labor between the two was once clear.

OpenAI provided increasingly powerful general-purpose models, while Harvey understood lawyers, legal documents, and law firm workflows, packaging these models into products usable by the legal industry. OpenAI owned the models, and Harvey owned the clients, placing them at different ends of the industrial chain.

However, cracks appeared in this upstream and downstream (upstream-downstream) relationship when foundational model companies began entering the application layer.

Fast forward to 2026, Jason Boehmig, co-founder of contract management software company Ironclad, joined OpenAI to develop legal industry products, directly collaborating with major law firms. From Harvey's perspective, its former ally was becoming a potential competitor.

If Harvey's core capabilities remained built on closed-source models like OpenAI's, any direct entry by upstream model providers like OpenAI into legal workflows would make it difficult for Harvey to establish truly independent technological boundaries and competitive barriers.

This dilemma is shared by most AI application companies.

While foundational models rapidly advance, the advantages built by application companies through prompts, knowledge bases, and workflows can be easily overshadowed by the next model update. Features requiring dedicated development yesterday may become standard capabilities of general-purpose models tomorrow. Model suppliers can erode the value of application companies by simply moving one step down the chain.

Beyond survival pressures, Harvey faced another issue—cost.

Harvey co-founder Gabe Pereyra stated in media interviews that from January to June 2026, the monthly token consumption on the Harvey platform surged from approximately 1 trillion to 14.5 trillion, a more than 14-fold increase in six months. Model calls were transforming from a technical procurement expense into a cost center affecting gross margins.

The explanation is straightforward.

Unlike simple chatbots, a legal agent might need to read hundreds of documents, continuously search for precedents, repeatedly call tools, and verify different conclusions. The more complex the task, the longer the call chain; the more widespread agents become, the faster token consumption grows.

This creates a hidden cost trap: clients and revenue belong to Harvey, but the escalating model costs associated with business growth continue to flow to suppliers. For a company valued at $11 billion aiming to become the legal AI infrastructure, this structure is unsustainable in the long run.

Kimi K3, representing open-source models, provided Harvey with an opportunity to 'escape' OpenAI.

In July 2026, Yue Zhi An Mian officially open-sourced Kimi K3, boasting 2.8 trillion parameters and supporting million-token contexts, with a focus on programming, knowledge work, and long-cycle agent tasks.

Independent AI evaluation platform Artificial Analysis conducted detailed tests on Kimi K3 using Harvey LAB-AA. In pass rate evaluations measuring the quality of legal deliverables, Kimi K3 ranked first with a 94.6% pass rate, surpassing Claude Fable 5. In stricter 'full task pass rate' evaluations, Kimi K3 led with 26.7%, nearly double that of the second-place Claude Fable 5.

Kimi K3 did not outperform the strongest closed-source models in all tests, but the gap has narrowed enough to influence enterprise selections.

For Harvey, the feasibility of replacing closed-source models with open-source alternatives became apparent for the first time: lawyers' work rarely involves answering isolated questions; it typically requires models to work continuously through large volumes of material, establishing factual relationships, identifying conflicts, assessing risks, and supplementing citations. A model's agent capabilities and ability to process complex knowledge are far more valuable than 'answering a single question correctly.'

Two indispensable factors further convinced Harvey to pursue post-training on Kimi K3.

The first is price.

Data from third-party evaluation platform Arena shows that Kimi K3 sits on the Pareto frontier of capability-cost, achieving approximately a +10.5% comprehensive net improvement at a task cost of about $0.62, less than one-fifth that of Claude Opus 5 (Max). Renowned Silicon Valley angel investor Jason Calacanis noted that some major clients of Anthropic and OpenAI reduced costs by 90% after deploying Kimi.

The second is permissions.

Unlike the limitations of tuning parameters and writing prompts for closed-source models, Harvey can modify Kimi K3 to suit its legal data, task environments, and evaluation systems; optimize inference paths based on task costs; and choose different infrastructure deployments. One sells intelligence to Harvey, while the other allows Harvey to transform intelligence into its own productive assets.

According to information disclosed in Harvey's official technical blog, the process was not a simple 'fine-tuning with legal materials.' Instead, Harvey constructed approximately 1,750 legal task environments, breaking down lawyers' factual judgments, citation requirements, risk ratings, revision suggestions, and delivery formats into scoreable expert standards. The model repeatedly completed full tasks, was scored against lawyer-developed criteria, and incorporated better work methods through reinforcement learning.

Test results from Legal Agent Bench showed that Tenet achieved a 'full task pass rate' of 19.7%, higher than Fable 5's 11.5% and GPT-5.6 Sol's 2.5%. In APEX Agents tests composed of enterprise legal tasks, Tenet scored 74.0%, also surpassing Fable 5's 67.4% and GPT-5.6 Sol's 60.3%. Currently, Tenet remains in the research preview stage but has already reached SOTA in Harvey LAB tests.

In simple terms, Kimi provides cutting-edge general capabilities, while Harvey ensures the model works like a lawyer: what to read first, how to search, how to judge evidence sufficiency, when to flag risks, and how to deliver results in the appropriate format.

Harvey's small step forward could represent a giant leap in reshaping the AI application market landscape.

A detail that should not be overlooked is that Tenet's post-training utilized only about 150 NVIDIA B300 GPUs and took just two months.

Using minimal computational power and cost, Harvey validated a replicable path for other industries, igniting the 'spark' for vertical AI companies to build their own models: open-weight models + industry expert data + task evaluation environments + post-training = proprietary vertical model capabilities.

For a long time, developing cutting-edge models was the domain of a few AI companies.

Training a general-purpose foundational model from scratch requires massive computational power, vast datasets, a complete research team, and continuously evolving infrastructure. Companies in legal, medical, financial, manufacturing, and other industries, despite possessing extensive professional knowledge, could not participate in model development itself. They typically integrated closed-source APIs and added prompts, knowledge bases, workflows, and product interfaces on top.

This model created the first generation of AI applications but also posed a problem: application companies owned the scenarios but not the full capabilities.

Harvey offers an alternative possibility.

Instead of retraining a general-purpose model from scratch, Harvey directly trained the professional capabilities required for legal work based on open-source models like Kimi K3. If a stronger open model emerges later, Harvey can migrate its task environments, evaluation criteria, and feedback data to the new foundation.

As model barriers lower, industry knowledge barriers will rise.

In the future, the companies best positioned to replicate Harvey Tenet's approach will not be all startups integrating large models; they will need four assets simultaneously: real business data, industry experts, quantifiable evaluation standards, and client workflows capable of forming feedback loops.

For example, medical companies need to know what diagnostic suggestions are missing, what evidence is insufficient to support conclusions; financial companies must define risk ratings, information citations, and compliance boundaries; manufacturing enterprises need to break down fault diagnoses, process experiences, and operational norms into tasks scoreable by models. These standards are difficult to obtain from public corpora and represent excellent opportunities for vertical companies to build barriers.

As Harvey officially stated, the next steps involve scaling computational power from 1,000 to 10,000 GPUs, conducting full-parameter fine-tuning, and exploring more open foundations. The goal is to enable every law firm to have its own proprietary model. Media outlets such as South Korea's Digital Today, Global Economic, and Japan's Livedoor have included the Harvey case in the industrial trend of 'vertical AI companies shedding reliance on cutting-edge model APIs.'

Predictably, AI application companies will no longer merely act as distributors of foundational model capabilities. A new commercial flywheel has formed: more clients lead to richer task environments; more expert feedback leads to more precise evaluation systems; stronger model capabilities widen the gap between products and general-purpose large models.

As more AI application companies turn to open-source models, they will shatter the entrenched belief that 'only closed-source top-tier large models can support commercial applications.' Open-source models will transition from 'low-cost alternatives' to the core of 'industrial digital infrastructure,' making 'strong open-source foundations + in-depth vertical post-training' the new paradigm for digital transformation across thousands of industries.

Harvey's story of building a proprietary model based on Kimi K3 may seem like a mere technical realignment, but it could spark an 'independence war' at the application layer.

In the second half of the large model race, the spotlight is shifting. AI applications that merely serve as shells may become relics of a bygone era, while countless decentralized, specialized vertical industry models, composed of 'Harveys,' will rise at unprecedented speed.

Meanwhile, Chinese open-source models will increasingly serve as the foundations for global AI application companies, a fact that closed-source giants cannot shake.