OpenAI-Backed Harvey Builds Legal AI Model on Kimi K3, Not OpenAI's Models

Harvey, the legal AI platform in talks to raise fresh capital at a reported $15.5 billion valuation, announced on August 20 that its first in-house AI model — Harvey Tenet — was post-trained on top of Kimi K3, the open-weight model released by Chinese lab Moonshot AI in July. The OpenAI Startup Fund is among Harvey's investors, yet its first proprietary model runs not on GPT but on a Chinese open-weight architecture, developed alongside US inference-infrastructure partner Fireworks AI.

The decision was driven by economics and competitive exposure, not ideology. Harvey processes more than 13 trillion tokens per month on behalf of more than 1,300 client organizations — law firms and corporate legal departments in 60-plus countries — and every one of those tokens has historically been billed by a third-party foundation-model provider. At those volumes, per-token API pricing from OpenAI, Anthropic, or Google is not a line item; it is a structural cost that the model vendors can reprice at will. The problem became more urgent when those same vendors began moving into Harvey's market. OpenAI has moved to compete directly in legal AI enterprise. Anthropic released legal-specific Claude plugins. Harvey's suppliers were becoming its competitors — and Harvey was paying for the ammunition.

Owning the model weights changes the calculus. Tenet runs at what Harvey describes as under one-quarter the per-token cost of comparable closed frontier models, with performance on legal-specific benchmarks that Harvey says exceeds both Fable 5 and GPT-5.6 Sol on the tasks that matter most to practicing lawyers. The company was explicit about the relationship between these two facts: co-optimizing for cost and quality during post-training, not just quality alone, is a deliberate architectural choice that affects every inference run going forward.

Read more: Moonshot AI Seeks Revenue Cut From Azure, AWS, and Google Cloud for Kimi K3

The Base Model Harvey Chose and Why It Fit

Kimi K3 is, by parameter count, the largest open-weight model currently available to developers. Moonshot released its full weights on July 27 under a Modified MIT license — with a commercial-agreement requirement for operators generating more than $20 million annually from K3-based services. At more than $350 million in annualized revenue, Harvey is firmly above that threshold, which means the company almost certainly negotiated or is negotiating a commercial agreement with Moonshot.

The model's architecture makes it particularly well-suited to legal work. K3 uses a mixture-of-experts design with 2.8 trillion total parameters and 896 separate expert subnetworks, of which only 16 activate for any given input token. The effective compute per token is closer to 104 billion parameters — which is what makes the economics viable at Harvey's scale. Running every parameter on every token at 2.8 trillion parameters is not commercially practical. Running 104 billion effective parameters through sparse activation is.

The key to K3's ability to sustain a one-million-token context window at commercially viable latency is an architecture Moonshot calls Kimi Delta Attention, or KDA. Standard transformer attention scales with the square of sequence length, making very long context windows computationally prohibitive. KDA interleaves full and linear attention layers in a 3:1 ratio — cheaper linear attention handles most of the computational load, while full attention is reserved for positions that need it most. The result is fast decoding at very long contexts with substantially reduced key-value cache memory. For legal tasks that routinely require reviewing thousands of pages of merger-agreement documents, due diligence data rooms, or litigation materials, a million-token window that actually runs at production latency is not a benchmark number; it is a functional requirement.

Harvey evaluated K3 on this basis against the available alternatives. Meta's Llama 4 is the other serious open-weight option in the frontier tier, but it lacks the long-context architecture that legal's most demanding tasks require. Alibaba's Qwen3.8-Max is a capable alternative at 2.4 trillion parameters but has not yet achieved the same breadth of independent evaluation that K3 has. On Harvey's own Legal Agent Benchmark, run in a controlled independent evaluation by Artificial Analysis, K3 scored 26.7% on the all-pass rate measure — nearly twice the next-best open-weight alternative — before any Harvey-specific post-training. That meant K3 arrived with more legal aptitude baked in than any other open-weight model Harvey could have chosen as a base.

How Harvey Trained Tenet: Asynchronous RL at Legal Scale

The post-training methodology Harvey used — described in detail in the company's August 20 technical blog post co-authored by Gabe Pereyra, Calvin Qi, and three other researchers — illustrates what it now takes to build a frontier-tier domain model without frontier-tier compute.

Harvey and Fireworks AI started from the K3 weights and applied asynchronous reinforcement learning using a method called Group-Sequence Policy Optimization, or GSPO. The core idea of GSPO is that for each training task, the system generates a group of independent agent attempts — different trajectories through the same legal problem — scores each against an expert-authored rubric, and learns from the relative performance within the group. This is more stable than training on single rollouts, because the learning signal reflects comparative quality rather than an absolute score that may be noisy.

The training environments were constructed by practicing attorneys, not drawn from public legal databases or existing case archives. Each environment consisted of a partner-level task instruction — a brief natural-language request written as if from a senior partner, typically around 50 words — paired with a simulated client matter containing a mix of key and peripheral documents, and a granular expert rubric of binary pass/fail criteria. On average, each task included 50 individual criteria, with the largest environments containing hundreds. A criterion might ask whether the agent correctly identified a specific risk in a fictional M&A target's environmental disclosures, whether it cited the relevant clause with the correct precision, or whether its reasoning chain was internally consistent from evidence to conclusion.

To achieve a passing grade on a task, an agent had to satisfy all criteria — not most. A single missed risk or inaccurate citation causes the entire task to fail. This all-pass structure is what Harvey means when it reports that Tenet "nearly doubles" the task-completion rate versus base K3: the standard for success is extremely high, and improvement at that standard is operationally significant.

The reinforcement learning used rank-64 LoRA adapters — low-rank matrix pairs inserted into all attention layers, feed-forward layers, and each of K3's 896 routed expert layers, adapting approximately 500,000 expert tensors in total. Full-parameter fine-tuning of a 2.8-trillion-parameter model requires far more GPU memory and training time than LoRA, which constrains the update to a low-dimensional subspace while leaving the bulk of the original weights intact. The result is a model that retains K3's general legal knowledge and reasoning capability while developing substantially improved behaviors for the specific patterns Harvey's training environments rewarded: thorough document traversal, precise citation, organized long-horizon work product, and efficient tool use that avoids burning unnecessary tokens.

The infrastructure challenge for training a large MoE model on RL deserves attention, because it is non-trivial. In reinforcement learning, the model that generates rollouts must be numerically close to the model that computes gradient updates. For a large MoE architecture running at finite precision — where routing decisions are made by a learned gating network, not deterministically — tiny numerical differences between the serving system and the training system can cause gradients to reflect a different policy than the one that actually generated the data. This is called off-policy error, and in severe cases it destabilizes training.

Fireworks AI solved this by building the trainer and rollout-serving deployments at the kernel level, aligning their numerical behavior. A token-in-token-out interface with router replay guarantees that the trainer recomputes log-probabilities using the same routing decisions that generated the original rollout. After each GSPO optimizer step, new weights are hot-loaded into the serving fleet without restarting it — so the GPU fleet generating rollouts is never idle, and the training loop stays continuous over the two-month run. The entire training process consumed approximately 150 Nvidia B300 GPUs, a figure that is roughly three orders of magnitude below what would be required to pre-train a model of comparable capability from scratch.

Benchmarks: What Harvey's Numbers Mean and What They Don't

Harvey Tenet's benchmark results are substantial by any reasonable measure, but the interpretive context matters for readers making technical or purchasing decisions.

The headline results come from Harvey's own Legal Agent Benchmark, or LAB, an open-source suite of more than 1,200 agentic legal tasks across 24 practice areas. Harvey designed and operates this benchmark; it is not a third-party standard. On the LAB hold-out set — tasks the model had not seen during training — Tenet completes nearly twice as many tasks as base K3, raising the all-pass rate by 9 percentage points in absolute terms. On the LAB: Contracts extension, which covers 500 tasks in in-house contracting work including drafting, review, and negotiation, Tenet ranks first and improves over K3 base by 2 percentage points in all-pass rate.

The distinction between "the model Harvey tested on a benchmark Harvey designed" and "an independently verified result" is one Harvey's own researchers acknowledged. The LAB evaluation framework includes methodological variations across different external evaluators — Vals runs Harvey's own benchmark with its own agent infrastructure; Artificial Analysis uses its stirrup harness and a Gemini 3.1 Pro grader; Harvey uses its own internal harness for its own scores. These variations can shift scores meaningfully.

The more independently credible data points come from evaluations Harvey did not control. Mercor, which runs the APEX Agents benchmark of professional-services agentic tasks authored by practicing attorneys, evaluated Tenet on its corporate lawyer subset and found it ranked first on the leaderboard — outperforming every other tested model. Mercor ran this evaluation with its own infrastructure and grading, without disclosing the tasks or task-level scores to Harvey before reporting. Crosby's Redline Bench, a multi-turn contract-negotiation benchmark graded against attorney-authored rubrics, also showed Tenet substantially outperforming the K3 base in evaluations Harvey ran using Crosby's standard configuration.

Where Tenet does not shine is worth stating clearly. On the PRBench hard subset — Scale AI's benchmark of high-stakes professional reasoning in law and finance — Tenet showed a non-statistically significant improvement over K3 base, moving from 36.0% to 36.8% criteria pass rate. Harvey acknowledged this explicitly. The improvement on LegalBench, the established academic benchmark of 162 legal reasoning tasks, was minimal, which Harvey also reported — but framed as acceptable because LegalBench tests static legal knowledge rather than agentic task completion, and Tenet was optimized for the latter.

The 51% hallucination rate that Artificial Analysis documented in independent testing of the base K3 model has not been addressed for Tenet specifically — Harvey published no hallucination-rate data for its post-trained model. For law firms making deployment decisions, this gap matters. A model that ranks first on agentic task completion but produces substantial rates of confident incorrectness in knowledge-retrieval settings creates downstream liability exposure.

The Broader Pattern: Kimi as the Open-Weight Base of Choice

Harvey is not an isolated case. The same week Harvey's announcement circulated in the legal AI community, analysis of the open-weight model landscape documented a consistent pattern: in much of 2026, the largest and highest-performing open-weight model available has been produced by a Chinese lab rather than a US one — and some US model releases above 100 billion parameters have been built on or influenced by artifacts from Chinese labs.

The pattern across sectors is consistent. Cursor's Composer 2 coding model, revealed by developer API inspection earlier in 2026, runs on a Kimi K2.5 base. Cognition's Devin integrated K3 weights into Devin Desktop and CLI after the weights were released in late July. Perplexity evaluated K2's performance immediately after its open-weight release and began post-training. At the infrastructure level, Chinese models peaked at 46.4% of routed enterprise tokens on OpenRouter in mid-2026, according to a CNBC report published in July — a figure that was 11% averaged over the prior twelve months. Open-source Chinese models price at roughly 60% to 90% below US frontier closed models, a differential that compounds at Harvey's 13-trillion-tokens-per-month scale.

The structural logic behind Kimi's selection as a post-training base across sectors is not that it outperforms every alternative on every benchmark. It is that K3 combines three properties that are individually available from US alternatives but not yet combined at the same tier: a very long context window at production-viable latency (enabled by KDA's linear-attention hybrid), an open-weight release that allows deep modification including LoRA over all expert tensors, and a pre-training data distribution that has demonstrated strong performance on domain-specific agent evaluations before any fine-tuning. That combination, at a price that reflects open-weight economics rather than closed-model API margins, is what Harvey, Cursor, and Cognition were each optimizing for when they chose Kimi as their starting point.

What This Means for Foundation-Model Economics in Vertical AI

Harvey Tenet's deepest significance is not that an OpenAI portfolio company chose a Chinese model — though that optic has drawn most of the commentary. The deeper significance is structural: the playbook Harvey followed is reproducible by any domain-specific AI company with sufficient expert data and approximately 150 high-end GPUs.

Pre-training a frontier model from scratch requires tens of thousands of GPUs running for months, capital commitments in the hundreds of millions to billions of dollars, and access to internet-scale training data. Harvey did not do that. Harvey used K3's pre-trained weights as a foundation — a representation of language, law, and reasoning already encoded at enormous cost by Moonshot — and applied domain-expert reinforcement learning to reshape those capabilities toward specific professional behaviors. The total compute was approximately 150 B300 GPUs over two months. The data came from practicing attorneys inventing fictional case scenarios and grading model outputs against professional standards.

The economic implication for foundation-model companies selling API access to specialized markets is significant. If a vertical AI company can achieve frontier-tier domain performance at sub-API-cost inference by post-training an open-weight base with expert data, the case for paying premium API prices narrows substantially. Harvey itself is explicit that Tenet is a bargaining chip: the model coexists with Harvey's Claude and GPT integrations rather than replacing them, but owning a cost-competitive alternative changes the terms under which Harvey negotiates with those providers.

Co-founder Gabe Pereyra, in his technical write-up accompanying the announcement, stated the longer-term goal plainly: Harvey intends to give law firms the capability to build their own proprietary models — fine-tuned on a firm's specific precedents, case archives, and institutional knowledge — on top of Tenet as a base. That would move the ownership of legal intelligence from Harvey's shared platform to individual firm infrastructure, creating proprietary models that competitors cannot replicate from public training data.

Harvey Tenet's Open Questions Before Production

The August 20 announcement is explicitly labeled a research preview, not a production release. Harvey described Tenet as demonstrating "promising initial results" rather than a deployment-ready system. The company says it plans to scale its compute from approximately 1,000 GPUs to 10,000 as it moves toward full-parameter fine-tuning — which would replace LoRA's efficient-but-bounded adaptation with direct modification of all parameters, typically improving task performance at the cost of substantially higher training compute.

Several questions remain materially open for law firms evaluating Tenet as a future platform component.

The hallucination question has not been resolved publicly. K3's 51% hallucination rate in base testing is documented by a credible independent evaluator. Harvey's post-training rewarded thorough citation and grounding, which should reduce this rate, but the company published no independent measurement of Tenet's hallucination behavior. Law firms exposed to malpractice liability for AI-generated work product need a credible answer to this question before production deployment.

The Moonshot commercial relationship creates a compliance layer that extends beyond Harvey's own terms of service. Kimi K3's Modified MIT license requires any operator generating more than $20 million in annual revenue from K3-based services to enter a commercial agreement with Moonshot AI — a Beijing-headquartered company subject to China's National Intelligence Law. Harvey's inference runs entirely on Harvey's own infrastructure; client matter data does not transit Moonshot's servers during Tenet operation. But the existence of a commercial relationship between Harvey and Moonshot, if it exists, creates a documentation chain that a law firm's own data security review will need to evaluate. Harvey has not publicly disclosed the terms of any such agreement.

The regulatory environment for open-weight Chinese AI in US enterprise is intensifying. Anthropic publicly alleged in February 2026 that Moonshot AI was among three Chinese labs that orchestrated industrial-scale distillation attacks on Claude using more than 24,000 fraudulent accounts. No enforcement action or Entity List designation against Moonshot has been publicly announced as of August 28. But the trajectory of US-China AI policy means that the compliance implications for US enterprises using Kimi K3-based products will require ongoing legal monitoring. The model weights themselves, once distributed under an open license, are not recalled by a subsequent designation — but the nature and terms of any ongoing commercial relationship between Harvey and Moonshot could be affected by changes in the regulatory landscape.

The next competitive milestone to watch is not another Kimi-based model. It is whether Harvey's investment in post-training infrastructure and training methodology produces generalist models with broader capability gains — and whether the 10,000-GPU scale-up the company has described enables full-parameter fine-tuning that closes the remaining gap with base K3 on benchmarks like PRBench where Tenet's current gains are not yet statistically significant. The broader question the industry is now running an experiment to answer is whether 150 GPUs of domain-expert RL is a template or a ceiling — and Harvey Tenet is the first large-scale legal test of that hypothesis.