On August 20, 2026, legal AI company Harvey released the Harvey Tenet Research Preview, publicly disclosing for the first time its interim results from post-training an open-weight model for specialized legal intelligence. Tenet is built on Moonshot AI’s Kimi K3 and was jointly post-trained by Harvey and Fireworks AI. Its objective is not simply to improve the accuracy of general legal question answering, but to strengthen the model’s ability to perform long-horizon, complex, multi-step legal work—including contract review, legal memoranda, due diligence, document analysis, and other forms of agentic legal work that require sustained tool use and the processing of large volumes of material.
Harvey defines two goals for the research with unusual clarity. First, it aims to build frontier-level legal intelligence on top of an open-weight model. Second, it seeks to establish a technical system through which law firms and professional-services organizations can train their own specialized models and own their own professional intelligence assets.
Tenet is therefore both a model experiment and a test of a new enterprise AI technology path: once an open foundation model becomes sufficiently capable, can an enterprise use task environments, expert evaluation criteria, an Agent Harness, and reinforcement learning to train the organization’s own methods of professional work back into the model?
Harvey Tenet Research Preview: Facts, Methods, and Results
Based on the information currently disclosed by Harvey and Fireworks, Tenet can be understood as a complete task-environment-driven post-training system, rather than conventional fine-tuning on legal corpora.
| Dimension | Publicly Disclosed Facts About Harvey Tenet | Technical Significance |
|---|---|---|
| Base model | Kimi K3 open-weight model | Continues training an open model that already has strong agentic and long-horizon capabilities |
| Training objective | Long-horizon agentic legal work | Optimizes the ability to complete legal work rather than merely answer legal questions |
| Training method | Asynchronous RL + GSPO | Learns work strategies from outcome feedback across complete task trajectories |
| Parameter updates | Rank-64 LoRA across attention, MLP, and routed experts | Adaptation extends beyond output layers and more broadly modifies model behavior and task-execution policy |
| Training environments | Approximately 1,750 agentic legal task environments | Each environment simulates a complete legal work assignment |
| Environment components | Partner instruction, client matter, expert rubric, tools, sandbox | The object of training is the complete work process |
| Data sources | Synthetic, public legal, and human expert data | Harvey explicitly states that no client data was used for this post-training effort |
| Training scale | More than 10,000 rollouts per epoch | Training is based on large numbers of complete agent trajectories |
| Compute | Approximately 150 NVIDIA B300 GPUs for roughly two months | This is already a substantial industry-scale RL post-training effort |
| LAB result | 19.7% all-pass vs. 10.8% for Kimi K3 | Nearly doubles the number of tasks passed in full |
| LAB Contracts | 11.3% vs. 9.3% for base K3 | Additional improvement on contract-related tasks |
| APEX Corporate Lawyer | Tenet: 74.0% | Demonstrates transfer to a third-party agent benchmark not included in training |
| Redline Bench | 55.5% vs. 49.3% for K3 | Improvement transfers across different Harnesses |
| Inference cost | $5.92 per LAB task vs. $5.62 for K3 | Capability gains were achieved with little increase in per-task inference cost |
Within these training environments, a typical task begins with only about 50 words of high-level partner instruction rather than a detailed prompt. The Agent must enter a client matter containing both relevant and distracting materials, search for files independently, read documents, establish case context, invoke tools, form judgments, and produce the final work product.
An expert rubric then decomposes partner-level review standards into atomic criteria covering facts, conclusions, citations, risk classifications, recommendations, formatting, and other dimensions. An average task contains roughly 50 criteria, while larger tasks may contain several hundred. A single rollout can exceed 1,000 interaction steps and consume hundreds of thousands of tokens.
Training uses GSPO. Multiple independent trajectories are generated for the same task and scored against the expert rubric. Group-relative advantages are then calculated so that the model receives feedback about which work strategies are more likely to produce high-quality outcomes.
The reward incorporates criteria satisfaction, completeness in resolving the underlying legal problem, and an additional reward for passing all criteria. It also applies some length constraints to encourage more efficient execution.
Harvey further emphasizes the importance of numerical consistency between the training and production inference environments. For reinforcement learning over extremely long trajectories, a training-serving numerical mismatch can be sufficient to destabilize training or even cause reward collapse.
The most direct result is that Tenet achieves a 19.7% all-pass rate on the LAB hold-out set, compared with 10.8% for the original Kimi K3. LAB Contracts improves from 9.3% to 11.3%.
Fireworks also reports that Tenet reaches 74.0% on Mercor APEX Agents Corporate Law, while its Redline Bench score increases from 49.3% for the base K3 to 55.5%. These benchmarks were not included in the training data, and Redline Bench even uses a different Harness from the one used during training. At minimum, this indicates that the learned behavioral capabilities exhibit some degree of transfer across tasks and execution environments.
Equally important, these improvements do not appear to materially damage the model’s existing legal knowledge. Harvey reports no significant degradation on LegalBench, CUAD, MAUD, APEX v1, or other professional-reasoning benchmarks.
This matters because one of the largest risks in industry-specific post-training is that strengthening a particular category of agent behavior may degrade the foundation model’s original general reasoning and domain-knowledge capabilities.
Three Initial Conclusions from the Tenet Research Preview
First, Tenet should be understood as Agent Policy Post-Training, not conventional legal-domain fine-tuning.
The primary object of optimization is not “how much legal knowledge the model knows.” It is how the model behaves after receiving an ambiguous professional assignment: how it searches, reads, plans, invokes tools, identifies issues, organizes evidence, produces deliverables, and controls reasoning cost.
What is being trained is the policy for completing the work.
Second, Tenet’s performance gains result from the joint optimization of the Model, Harness, Environment, and Evaluation system. They should not be reduced to a claim that “fine-tuning Kimi K3 made it outperform Fable 5 and GPT-5.6 Sol.”
Harvey’s own experiments demonstrate this well. Under the official Mercor setup, Kimi K3 scores 58.8%. Simply switching to Harvey’s filesystem/Bash Harness raises the same base model to 67.5%.
Likewise, changing the Harness increases Fable 5’s score by 7.7 percentage points, while GPT-5.6 Sol declines by 4.2 percentage points under the same kind of change.
What an Agent benchmark measures is therefore no longer the Model alone. It measures:
Model × Harness × Tool × Context × Environment
Third, the new strategic value of open models does not primarily come from “cheaper APIs.”
Kimi K3 is a Mixture-of-Experts model with 2.8 trillion total parameters and 104 billion activated parameters, using 896 experts and activating 16 experts per token, while supporting a context window of approximately one million tokens.
More importantly, Moonshot provides open weights. That gives Harvey the ability to modify the model parameters further and train its own legal task-execution strategies into the model.
For enterprises, Open Weight therefore represents a transition from merely “renting intelligence” toward processing, adapting, and owning proprietary professional intelligence.
What Harvey Is Actually Building Is a Production System for Industry-Specific Proprietary Intelligence
Abstracted beyond the legal industry, the engineering architecture behind Tenet can be represented as:
Frontier Open Model → Domain Task Environment → Agent Harness → Expert Rubric → Trajectory Generation → Reward → Post-Training → Independent Eval → Production Harness → Continuous Improvement
The order matters.
An enterprise begins by defining real work—not by selecting a training algorithm.
Harvey converts the way law-firm partners actually delegate work into task instructions, converts matter files into environments, converts professional review standards into rubrics, and converts the model’s entire process of completing the assignment into trajectories.
Only after those elements exist does reinforcement learning have something meaningful to optimize.
For this reason, an enterprise seeking to replicate the Tenet approach should not begin by building a GPU cluster. Its first assets should be a Task Environment Library and an Evaluation System.
Representative business work must first be expressed as realistic tasks:
What input materials are provided?
Which tools may the model use?
What are the possible steps required to complete the work?
What constitutes the final deliverable?
How is correctness determined?
Which errors are unacceptable?
Only after this foundation has been established should the enterprise build the Harness.
The Agent must have the execution capabilities required for real work, including access to filesystems, search, databases, code execution, browsers, enterprise APIs, business systems, and state management.
Harvey’s experiments already demonstrate that even a highly capable model cannot fully express its capabilities if the Harness itself is poorly matched to the task.
The third stage is to establish a baseline.
At minimum, the enterprise should compare several frontier closed models with multiple open-weight models and record metrics such as:
Quality, All-Pass Rate, Cost, Latency, Tool Success, Human Review, and Failure Distribution.
Only after determining that the remaining problems arise primarily from systematic deficiencies in model behavior—rather than from Prompt, RAG, Context, or Tool design—does post-training become justified.
The fourth stage is SFT, Distillation, or RL.
The training objective should not be a vague instruction such as “answer more professionally.” It should target clearly identified failure modes, such as:
insufficient search coverage, premature termination, inaccurate citations, failure to invoke required tools, inadequate verification, weak task decomposition, or unstable output structure.
The Reward must directly correspond to business acceptance criteria.
Finally, the organization should rerun a completely independent hold-out benchmark.
Only if the new model improves not only on its training tasks but also on unseen tasks, different Harnesses, or real production assignments—and does so without serious regression—can the organization claim that it has created a genuinely generalizable capability.
Why the Harness May Be as Important as Post-Training Itself
One of the most important data sequences in the Tenet research for enterprise AI engineering teams is not 74.0%, but:
58.8% → 67.5% → 74.0%
The 58.8% score represents base Kimi K3 under the Mercor Harness.
After switching to the Harvey Harness, the same base model reaches 67.5%.
After Tenet post-training, it reaches 74.0%.
Strictly speaking, the three results do not constitute a perfectly controlled experiment. Nevertheless, they reveal an important property of Agent systems: a substantial proportion of what appears to be “insufficient model capability” may actually result from a mismatch between the model and its execution environment.
This changes the technical order of operations for enterprise AI.
When an Agent repeatedly fails, the correct diagnostic process should not immediately point to fine-tuning.
The first questions should be:
Is the Context correct?
Are the Tools sufficient?
Are tool semantics clear to the model?
Is task decomposition appropriate?
Does the Harness expose primitives suited to the work?
Does the Agent maintain the necessary state?
Does the Evaluation system actually measure the intended business outcome?
Harvey’s experiment in M&A due diligence reinforces the same conclusion.
A single due-diligence assignment may require traversing as much as approximately 80 million tokens of material, far beyond what any single context window can effectively process.
Harvey and Baseten therefore used a Recursive Language Model Harness, allowing a root agent to operate over the full data room through a REPL environment while delegating research tasks to sub-agents with independent contexts.
Changing the Harness alone enabled GLM-5.2 to reach a 46.1% criteria pass rate. Further self-distillation post-training within that Harness increased the score to 60.1%.
Model, Agent Harness, and post-training can therefore no longer be designed independently.
A more appropriate principle is:
train the model inside the environment where it will work.
Tenet Redefines What “AI Data” Enterprises Should Accumulate
Historically, when enterprises discussed AI data assets, they usually began with documents, knowledge bases, databases, and vector stores.
Tenet suggests a different answer for the next phase of Agent AI:
The genuinely scarce data is data about how the enterprise performs its work.
This includes:
Real Tasks, Input Matter, Expert Rubrics, Tool Environments, Successful Trajectories, Failure Trajectories, Human Corrections, and Final Outcomes.
Documents describe what the enterprise knows.
These new forms of data describe how the enterprise converts knowledge into outcomes.
The Expert Rubric is particularly important.
Without a rubric, an enterprise can observe what a model generated, but it cannot consistently calculate how far the result is from acceptable professional work.
Without stable evaluation, systematic Eval becomes impossible—and without Eval, reliable Reward cannot be generated.
The foundation that enables Harvey to conduct reinforcement learning is therefore not simply its access to Kimi K3. It is the fact that Harvey had already begun engineering the evaluation standards of professional legal work.
This is also why, as enterprise AI enters the Agent era, Eval should precede Post-Training.
Only after an enterprise establishes measurable business tasks can it systematically generate training data and reinforcement-learning environments.
The Cost Model Should Also Shift from Token Price to Accepted Outcome
Another important Tenet result concerns cost.
Base Kimi K3 costs approximately $5.62 per LAB task, while Tenet costs approximately $5.92 per task. Inference cost therefore remains broadly comparable, while the all-pass rate increases from 10.8% to 19.7%.
Harvey explicitly uses reward shaping to encourage more efficient tool use and reasoning trajectories during training. The model is therefore learning not only how to improve quality, but also how to reduce unnecessary computation.
Enterprises should consequently stop defining AI cost primarily as Dollars per Million Tokens.
A more meaningful measure is:
Cost per Accepted Outcome = Model Cost + Tool Cost + Retry Cost + Human Review Cost + Correction Cost + Failure Cost
The economic value of professional work ultimately depends on how many tasks reach the standard acceptable to business professionals—not on how many tokens the model generates.
Viewed from this perspective, the commercial value of post-training becomes much clearer.
Even if the cost of an individual model invocation does not decline significantly, higher first-pass completion rates, reduced human review, fewer retries, and shorter trajectories can still materially reduce the unit economics of the overall business process.
Tenet Does Not Prove That Every Enterprise Should Train Its Own Model
Harvey’s approach also has clear barriers to entry.
Approximately 1,750 task environments, more than ten thousand long-horizon rollouts, around 150 B300 GPUs, and roughly two months of training resources make this scale of reinforcement learning far beyond what an ordinary enterprise should undertake immediately after assembling a few datasets.
Organizations that are genuinely well suited to this path will usually satisfy several conditions simultaneously:
The task occurs repeatedly and carries substantial economic value.
Work quality can be evaluated explicitly.
The organization possesses significant expert knowledge and experience.
The Agent can already execute most of the workflow through tools.
Foundation-model errors exhibit stable and identifiable patterns.
Business volume is large enough to amortize training and maintenance investment.
When these conditions are not yet present, the more appropriate technical path remains:
Model Routing, Prompt Engineering, RAG, Tool Engineering, Workflow, Harness, and Eval.
Post-training becomes a high-value instrument only after these methods begin to reach a performance plateau and the organization can accurately identify the remaining failure patterns.
Enterprise AI strategy should therefore not be framed as a binary choice between “using APIs” and “building your own model.”
It is better understood as a continuous capability ladder:
Frontier API → Agent Harness → Domain Eval → Open-Weight Model → SFT/Distillation → RL Post-Training → Proprietary Intelligence
What Harvey Tenet Really Shows Is That Professional Intelligence Can Begin to Be Manufactured
Tenet remains a Research Preview.
Its 19.7% LAB all-pass rate also demonstrates that complex legal Agents remain far from being able to complete professional work fully autonomously.
LAB applies a particularly strict standard: every evaluation criterion must pass. If a task contains even one critical deficiency, it cannot be counted as successful. In this sense, the metric is much closer to the way professional work is actually accepted in practice than many conventional benchmarks.
But Tenet has already demonstrated something more important.
Once open foundation models enter the frontier capability range, professional-services organizations are no longer limited to building external applications around those models.
They can begin to train their own methods of working back into the model itself.
In the future, enterprises in legal services, finance, auditing, consulting, manufacturing, energy, and other sectors with high-value professional workflows may increasingly adopt a two-layer intelligence architecture.
The lower layer will consist of frontier open-weight models such as Kimi K3, Qwen, and DeepSeek, providing general intelligence.
The upper layer will consist of a Proprietary Intelligence Layer built from the enterprise’s own:
Task Environments, Harnesses, Tools, Knowledge, Rubrics, Memory, and Post-Training.
Foundation models provide world knowledge and general reasoning capabilities.
The competitive advantage an enterprise actually owns lies elsewhere:
It knows how its work should be performed, how to determine whether that work has been performed well, and how to convert that experience into training signals that machines can learn from.
The significance of Harvey Tenet is therefore not merely that “a Kimi K3 post-trained for legal work outperformed Fable 5 and GPT-5.6 Sol on certain legal benchmarks.”
A more accurate conclusion is:
Enterprises are beginning to acquire an engineering pathway for converting business work, expert experience, and evaluation systems into proprietary model capabilities.
That may ultimately prove to be a far more important long-term consequence of frontier open-weight models entering the enterprise market than lower inference costs.
Related Topic
Understanding a Law Firm through Study
The True Moat of Enterprise AI: From Data Readiness to Scenario-Based Closed-Loop Execution
LLMs Enter Enterprise Core Systems — The Real Question Is No Longer "Is the Model Strong Enough?"
Data Intelligence: Laying the Foundation for Enterprise AI
Enterprise AI Adoption: A Paradigm Shift from "Buying Models" to "Buying Business Outcomes"
Key Considerations for Fine-Tuning Generative AI Models
The Truth About Enterprise AI Deployment: Why 90% of Projects Never Make It Past the Demo Stage