Woofun AI reports that Harvey, the legal artificial intelligence unicorn backed by OpenAI, has officially launched Tenet, marking a decisive strategic pivot toward open-weight foundation models. This inaugural in-house model is built upon Kimi K3, an open-weight architecture released by Moonshot AI in July, signaling a departure from the company's previous reliance on closed-source APIs. The development team is led by founder Winston Weinberg, a former litigation attorney at O'Melveny & Myers, and Gabe Pereyra, who brings extensive large model research experience from Google Brain and Meta Platforms.

This collaboration represents a significant structural shift in how vertical AI firms approach proprietary model development, moving from passive API consumption to active weight customization. The launch of Tenet underscores a growing industry trend where specialized enterprises leverage open-source foundations to embed domain-specific knowledge directly into model weights, rather than relying solely on prompt engineering or external tooling.

By integrating Kimi K3, Harvey aims to create a more controllable and cost-effective infrastructure for handling complex legal tasks, while maintaining the high performance standards required by its global client base. This move also reflects a broader maturation of the open-weight ecosystem, where models like Kimi K3 have reached a level of sophistication that makes them viable candidates for enterprise-grade fine-tuning.

The decision to adopt a Chinese-developed open model highlights the increasing globalization of AI infrastructure and the diminishing relevance of geographic boundaries in model development. As Harvey transitions from a pure API-dependent service to a hybrid model with proprietary weights, it sets a precedent for other vertical AI companies seeking greater autonomy over their core technology stack. The success of Tenet will likely influence how other high-stakes industries, such as finance and healthcare, approach their own model customization strategies in the coming years.

Harvey was founded in August 2022 and has since achieved a valuation of $11 billion, establishing itself as the most highly valued legal AI firm in the world. The company's investor roster includes prominent venture capital firms such as Sequoia Capital, Andreessen Horowitz, and GIC, reflecting strong institutional confidence in its business model. Prior to the launch of Tenet, Harvey's technology stack was predominantly reliant on closed-source models, which were considered the standard in professional services industries due to their superior performance on complex reasoning tasks.

OpenAI was an early investor in Harvey, and the two entities collaborated on training a specialized case law model using approximately 10 billion tokens of U.S. case law data. In customer case studies released by OpenAI at the time, Gabe Pereyra noted that Harvey evaluated various options before deciding to rely solely on OpenAI for this custom training effort, highlighting the initial preference for closed-source solutions. This reliance on external APIs allowed Harvey to access some of the strongest models globally without the burden of maintaining its own infrastructure.

However, this approach also limited Harvey's ability to deeply customize the model's behavior to fit specific legal workflows and evaluation criteria. The company's rapid growth, driven by its ability to deliver high-quality legal assistance, created a demand for more granular control over model outputs. As Harvey expanded its client base, it became evident that closed-source models, while powerful, could not be fully aligned with the nuanced requirements of legal practice. The decision to develop Tenet was therefore not just a technical upgrade but a strategic necessity to maintain competitive advantage in a rapidly evolving market. By building its own model, Harvey can now integrate proprietary data and expert feedback directly into the training process, creating a more tailored and effective solution for its users.

This shift also reduces dependency on third-party providers, mitigating risks associated with API changes, pricing fluctuations, and service disruptions. The $11 billion valuation underscores the market's belief in Harvey's ability to scale this new model-centric approach across the global legal industry.

The transition from closed APIs to open weights represents a fundamental paradigm shift in how AI capabilities are developed and deployed. In the past, model companies completed pre-training and post-training before packaging their capabilities as interfaces, leaving downstream companies with limited ability to alter the model itself. Users could modify prompts, perform searches, and build agents, but the core weights remained fixed and inaccessible. This assumption held true when closed-source APIs were dominant, as the capabilities of a model were largely determined by the provider.

However, open weights break this assumption by allowing companies to further enrich the weights with their own data and expert feedback. This enables the incorporation of real-world tasks and evaluation standards directly into the training process, creating a more customized and effective model. The impact of this shift is evident in benchmark performance, where two models that both score 90 points at release can diverge significantly after continued training.

One model might score 91 points, while another reaches 97 points, depending on the quality and relevance of the additional training data. This divergence highlights the importance of having weights worth training on, as the potential for improvement is directly tied to the foundational capabilities of the base model. In recent years, there has always been a gap between open models and the most advanced closed-source models, particularly in professional scenarios where errors carry high costs.

Law is one of the fields with the highest requirements, and in the past, only the strongest closed-source models could truly compete in companies like Harvey's core business areas. Open weights are certainly easier to control and modify, but if they cannot handle even basic professional tasks, further training becomes meaningless. The emergence of models like Kimi K3 has helped overcome this long-standing barrier, providing a solid foundation for vertical AI firms to build upon.

This shift also changes the economic dynamics of AI development, as companies can now invest in customizing models rather than paying for API usage. The ability to retain specialized capabilities in proprietary weights offers a significant competitive advantage, as it allows firms to create unique value propositions that are difficult for competitors to replicate. As more companies adopt open-weight models, the industry is likely to see a greater emphasis on post-training and customization, rather than just pre-training and deployment.

Kimi K3's technical specifications played a crucial role in Harvey's decision to adopt it as the foundation for Tenet. With 2.8 trillion total parameters and support for 1 million token contexts, Kimi K3 possesses solid foundational capabilities for long-term tasks, tool usage, and complex inference. These features are essential for legal applications, which often involve processing large volumes of documents and executing hundreds or even thousands of operations in succession. The model's ability to maintain consistency throughout lengthy processes is particularly important, as legal tasks require a high degree of accuracy and attention to detail.

For Harvey, open-weight models were no longer just a cheaper, more controllable alternative; they entered the realm where they could be seriously evaluated alongside cutting-edge models. This led them to consider a question that had rarely been discussed separately: just how easy is it to further train these weights? The answer to this question depends on several factors, including the model's architectural design, the quality of the training data, and the expertise of the development team. Kimi K3's mixed expert structure, while complex, offers significant advantages in terms of efficiency and performance.

However, it also presents engineering challenges, such as ensuring that the same token is assigned to the same expert during training and execution. Harvey addressed these challenges by aligning numerical values at the kernel level between the trainer and execution environment, allowing for more accurate and stable training. The model's support for 1 million token contexts also enables it to handle long-trajectory tasks, which are common in legal practice.

This capability is critical for tasks such as M&A due diligence, where the model must search through hundreds or thousands of documents to identify relevant clauses and assess risks. By leveraging Kimi K3's technical strengths, Harvey was able to create a model that is not only more customizable but also more effective at handling complex legal tasks. The success of Tenet demonstrates that open-weight models can now compete with closed-source alternatives in high-stakes professional environments, provided that they are trained with high-quality data and expert feedback.

Precedents in other industries illustrate the growing trend of adopting Kimi models for specialized applications. Cursor, a popular coding assistant, did not choose a model based on common programming agent rankings when training Composer 2. Instead, it reasoned that agent capabilities and long-term task performance could change significantly during reinforcement learning, so pre-training rankings might not predict final outcomes. Rather than focusing on ranking scores, Cursor placed more emphasis on a model's programming knowledge, state tracking ability, confusion level when working with internal codebases, and efficiency within its own infrastructure.

The chosen foundation was ultimately the Kimi K2.5 model, which provided a solid base for further customization. Cognition, the company behind Devin Desktop, reached a similar conclusion. After Kimi K3's weights were made available, Cognition quickly integrated the model into Devin Desktop and its CLI, where it achieved 58.2% accuracy and 63.6% pass rate on FrontierCode 1.1 Extended. These tests evaluated real-world engineering tasks, taking into account whether the code could be integrated into the main codebase and its quality.

Cognition also noted that Kimi K3 performed particularly well at reproducing bugs and managing its own runtime environment—two areas that are especially prone to failure in long-term programming tasks. The company subsequently conducted post-training on these weights as well, further enhancing the model's capabilities. These examples demonstrate that Kimi models are not just suitable for general-purpose tasks but can also be effectively customized for specialized domains. The success of Cursor and Cognition in leveraging Kimi models for programming tasks suggests that other industries, such as law, finance, and healthcare, can similarly benefit from open-weight customization.

By focusing on domain-specific requirements and evaluation criteria, companies can create models that are more effective and efficient than those built on closed-source APIs. This trend is likely to accelerate as more open-weight models reach a level of sophistication that makes them viable candidates for enterprise-grade applications. The ability to customize models for specific use cases offers a significant competitive advantage, as it allows companies to create unique value propositions that are difficult for competitors to replicate.

Woofun AI data shows that Harvey's rapid growth provided the company with a wealth of data and insights that are difficult for foundation model companies to obtain on their own. By March this year, Harvey was serving around 1,300 organizations and over 100,000 lawyers. Five months later, its client base nearly doubled, covering 70 countries, with more than three-quarters of AmLaw 100 law firms using Harvey. This expansion allowed Harvey to accumulate a vast amount of legal data, including client interactions, task assignments, and evaluation criteria.

The company understood how lawyers assign tasks and where partners looking for errors would focus their attention. This experience was later compiled into the Legal Agent Benchmark (LAB), which includes over 1,200 long-term legal tasks across 24 practice areas. Each task description is typically only about 50 words long, after which the model enters a closed client matter environment filled with relevant documents and numerous irrelevant materials. It must search, read, analyze, and submit a work product that can be verified item by item.

There are usually around 50 evaluation criteria per task, with complex tasks having hundreds. Whether all facts were covered, whether conclusions hold up, whether citations match the original text, and whether risk levels and recommendations are reasonable—all are broken down into separate evaluable items. A single task can last over 1,000 rounds, consuming hundreds of thousands of tokens. This benchmark provides a rigorous testing ground for evaluating the performance of legal AI models, ensuring that they meet the high standards required by the legal profession.

The creation of LAB also allowed Harvey to break down tacit knowledge, such as what clients truly worry about and which seemingly minor clauses cannot be overlooked, into measurable and comparable feedback. This approach enables each completion cycle of the model to generate actionable insights that can be used to improve its performance. By incorporating these tasks and criteria into the training process, Harvey was able to create a model that is not only more accurate but also more aligned with the practical needs of legal practitioners.

The success of LAB demonstrates the value of domain-specific benchmarks in driving the development of specialized AI models.

The training methodology employed by Harvey for Tenet involved a combination of advanced techniques and substantial hardware infrastructure. Harvey prepared approximately 1,750 legal agent task environments, each equipped with materials and tools necessary for completing legal tasks. Once the model enters the workspace, it must read documents, use tools, and submit results. The system then evaluates the entire workflow against standards set by lawyers to determine how much of the requirements were met and to what extent legal issues were resolved. If all key requirements are satisfied, an additional reward is given.

Each training cycle generates over 10,000 task trajectories, providing a rich dataset for optimizing the model's performance. Harvey uses the group sequence optimization method, known as GSPO, to enable the model to generate multiple approaches to the same task and update weights based on differences in quality among the results. If two trajectories score too similarly, the system re-evaluates them to reduce the impact of evaluation noise on training. The entire process takes about two months and involves around 150 NVIDIA B300 GPUs.

Pre-training a cutting-edge model usually requires tens of thousands of GPUs, but Harvey's computational investment is on a completely different scale, reflecting the efficiency of post-training techniques. Harvey employs rank-64 LoRA, covering Kimi K3's attention layers, feedforward networks, and routing expert weights, involving approximately 500,000 expert tensors. Long-trajectory training also presents an engineering challenge, as legal tasks may span hundreds of rounds, and the weights on the training side may have already been updated by the time the model generates a task trajectory.

To address this, Harvey and its suppliers aligned the numerical values at the kernel level between the trainer and execution environment. After each weight update, it is directly hot-loaded into the execution environment without needing to pause the task repeatedly. The system also records the expert routing results corresponding to each token, allowing the trainer to follow the path taken by the model when generating that trajectory as closely as possible. This meticulous approach ensures that the training process is both stable and efficient, enabling Harvey to create a model that is highly optimized for legal tasks.

Performance results on internal and external benchmarks demonstrate the effectiveness of Harvey's training methodology. According to Harvey's internal results, on the untrained LAB holdout set, Tenet completed nearly twice as many tasks as the original Kimi K3, with the all-pass rate rising from around 11% to 19.7%, an increase of nearly 9 percentage points. In LAB Contracts, a dataset specifically designed to test contract drafting, review, and negotiation, the number of completed tasks increased by about 20%, with the all-pass rate reaching 11.3%, an increase of about 2 percentage points.

According to Harvey's own rankings, Tenet ranked first in LAB Contracts and second in the full LAB set. It was then tested in two external datasets: Mercor's APEX Agents Corporate Law and Crosby's Redline Bench. The former tests long-term professional tasks in a simulated work environment, while the latter requires the model to continuously revise contracts and be scored according to lawyer-defined standards. Despite never having been exposed to the training data for these datasets, Tenet still performed significantly better than the original Kimi K3, indicating that the capabilities developed in Harvey's task environment could be transferred to new legal tasks.

However, another issue is that specialized post-training often makes models better at certain types of work while causing them to lose some of their original knowledge and reasoning abilities. On LegalBench, CUAD, and MAUD, Tenet showed no significant decline. In the challenging subset of the Scale Professional Reasoning Benchmark, its score even increased slightly from 36.0% to 36.8%. At least from these results, it appears that Kimi K3 did not suffer significant capability loss after becoming stronger in legal tasks.

Harvey's reward design favors shorter task trajectories when results are of similar quality, as legal agents consume more tokens and incur longer waiting times for every additional document read or tool call. After training, Tenet reduced some unnecessary steps, improving task quality while keeping costs relatively stable. What changed the most after this post-training was how Kimi K3 handled complex legal tasks. It became better at organizing steps, using tools, and ensuring that key requirements in tasks were not overlooked.

These results highlight the potential of open-weight models to achieve high performance in specialized domains without sacrificing general capabilities.

The strategic implications of Harvey's approach extend beyond the legal industry, influencing how other vertical AI firms might develop their own models. As early as 2025, Harvey began adopting a multi-model strategy and continues to integrate new models from OpenAI and Anthropic this year. Tenet marks the first time that this system has included an open-weight model trained and controlled by Harvey itself. What's needed is a model whose foundation is already robust, as Harvey's legal tasks often involve processing large volumes of documents, executing hundreds or even thousands of operations in succession, and maintaining consistency throughout the lengthy process.

Post-training can further shape how the model handles legal work, but it's difficult to add capabilities that the foundation model originally lacked. Another requirement is control. Harvey can call upon GPT and Claude via APIs, but it cannot incorporate the 1,750 legal task environments and lawyers' evaluation standards into these closed-source models. Open weights allow it to decide how to train and design rewards, as well as retain the specialized capabilities in its own weights.

However, even after obtaining the weights, it's necessary to assess whether these models are suitable for further training. Whether long-trajectory reinforcement learning can run stably, whether the mixed expert routing can be accurately replicated, whether improved professional capabilities won't come at the cost of losing existing knowledge, and whether inference costs remain within acceptable levels for practical use—all these need to be tested firsthand. After two months of operation, Kimi K3 provided an answer to these questions.

The training process ran stably, legal task capabilities improved significantly, no notable capability loss was observed, and costs didn't spiral out of control despite the improvements. For Harvey, these results are more important than the concept of "open weights" itself. The fact that weights can be downloaded only indicates that a company is qualified to continue training. It is the amount of capability retained after training that determines whether investing additional computational resources, data, and expert time is worthwhile.

This strategic shift towards open-weight customization offers a new pathway for vertical AI firms to achieve greater autonomy and competitiveness in their respective markets.

The new market dynamic for vertical AI training on open weights is reshaping the landscape of professional services. Harvey states that its goal is to enable law firms to train their own proprietary models on open weights and retain the capabilities resulting from that training. Tenet represents not so much a model as much as a methodology, demonstrating how companies can take an existing foundation model and continue training it, integrating their industry-specific knowledge into the weights to create a model better suited to their business needs. Such demand was difficult to sustain as a distinct market in the past, when general-purpose model capabilities were insufficient. Vertical companies could hardly compensate for weaknesses in long-term reasoning, tool usage, and complex task handling merely through post-training, even if doing so required expensive custom pre-training, which most companies found unnecessary.

However, with the emergence of sophisticated open-weight models like Kimi, this barrier has been lowered. Cursor developed Composer 2 in the field of programming, Cognition continued training the model used by Devin on the same weights, and Harvey created Tenet in the legal domain. Although programming and law are quite different fields, both started by building upon Kimi's already trained general capabilities and then adding their own specialized work processes. Tenet also makes the barriers to entering this space more concrete.

With two months, around 150 GPUs, and evaluation standards defined by practicing lawyers, a company can advance a general-purpose weight set to the forefront of its industry. The computational threshold has dropped by an order of magnitude, while the more difficult aspects lie elsewhere: determining precisely what level of completion constitutes a finished professional task. Software development and law are just two examples that have already emerged. Finance, consulting, pharmaceuticals, accounting, and numerous other professional service industries all have their own data, workflows, and expert judgments.

These industries constitute knowledge-based work markets worth trillions of dollars. Even the leading companies in these fields may not pre-train their own foundation models, but they increasingly have the resources to train their own models on mature weight sets. If this trend continues, Kimi's market will no longer be determined solely by how many people use it. What's more noteworthy going forward is how many companies, following Harvey, Cursor, and Devin, will choose to train their own models on Kimi.