An American legal-AI company used a Chinese open-weight foundation model, U.S. training infrastructure and expert legal environments to build a specialist that beats its base model on demanding legal work. The achievement is real. So are the caveats.
Reporting status: Source-audited analysis; no hands-on Tenet test is claimed.
Evidence cutoff: August 21, 2026.
Verdict: Harvey Tenet is the clearest evidence yet that frontier pretraining is becoming industrial raw material. The defensible value is moving into domain environments, expert feedback, post-training, agent harnesses and deployment economics. Its benchmark results are impressive, but they remain a research preview built largely on company-run evaluations.
There is a revealing contradiction at the centre of Harvey Tenet.
Harvey is one of America’s most prominent legal-AI companies. Tenet, its first model post-trained specifically for legal work, begins with Kimi K3: a 2.8-trillion-parameter foundation model created by Beijing-based Moonshot AI and released with downloadable weights. Harvey then combined public and synthetic legal data, human-expert data, realistic legal assignments, a carefully designed reward system and roughly two months of reinforcement learning on U.S. infrastructure.
The result, according to Harvey’s unusually detailed research preview, nearly doubled Kimi K3’s strict task-completion rate on the company’s Legal Agent Benchmark while keeping inference cost roughly stable. Tenet placed second on the broader LAB evaluation and first on LAB Contracts in Harvey’s reported snapshot. It also improved on the base model across two outside agent benchmarks.
That is worth celebrating. It is also more interesting than another model leaderboard win.
Tenet shows what happens when a vertical software company no longer has to build frontier intelligence from scratch or rent every important capability from a closed API. Harvey treated Kimi K3 as a starting material, then added the part that mattered for its users: a working definition of excellent legal work.
Tenet at a glance
| Item | What Harvey disclosed | Why it matters |
|---|---|---|
| Foundation | Moonshot AI’s Kimi K3 open-weight model | Harvey started with frontier-scale general capability rather than pretraining from zero |
| Training partners | Fireworks AI for Tenet; additional specialist work with Baseten, Applied Compute and Engram | The project is a stack of model, infrastructure and domain-research companies |
| Data | Public legal data, synthetic data and human-expert data; Harvey says no customer data was used | Expert supervision and realistic work environments are central to the result |
| Method | Asynchronous reinforcement learning using GSPO and rank-64 LoRA | Tenet changes model behaviour, not only prompts around the model |
| Scale | About 1,750 environments, more than 10,000 rollouts per epoch, roughly 150 NVIDIA B300 GPUs for two months | This is substantial frontier-model engineering, even though it is far smaller than pretraining K3 |
| Headline result | 19.7% all-pass on LAB and 11.3% on LAB Contracts | Strong under a deliberately unforgiving metric, but far from solved legal autonomy |
| Status | Research preview connected to the Harvey II product story | Harvey has not published a reproducible Tenet checkpoint, API or complete training corpus |
What Harvey actually built
Calling Tenet a “fine-tune” is technically defensible and editorially misleading. It hides most of the work.
The base model is Kimi K3, a sparse mixture-of-experts system with 2.8 trillion total parameters, about 104 billion activated for each token, 896 routed experts and a configured context window of 1,048,576 tokens. Moonshot’s model card describes native multimodality and long-horizon agentic work as core design goals. The full checkpoint is downloadable, although at roughly 1.56 TB it is emphatically not a laptop model; Kingy.ai’s Kimi K3 weights and hardware guide details what the release contains and the infrastructure it requires.
Kimi’s custom licence broadly permits use, modification and distribution, but it is not a standard open-source licence. It adds conditions for large model-as-a-service businesses and very large commercial products, with exceptions including internal use and access through certified inference partners. “Open-weight under the Kimi K3 License” is therefore more precise than the casual “open source” label. The licence text is public.
Harvey’s contribution sits above that foundation.
It built legal-task environments modelled on how work moves through a major firm. A task begins with a short instruction written like a partner’s request to an associate. The agent enters a sandbox containing a fictional client matter, relevant and irrelevant documents, search and drafting tools, and a required work product. An expert rubric then decomposes acceptable work into binary criteria covering facts, conclusions, citations, deadlines, severity ratings, recommendations and formatting.
The distinction matters. A conventional legal benchmark may ask a model to identify a clause or answer a doctrinal question. Harvey’s open-source LAB project asks an agent to find the right files, form a plan, use tools, reconcile scattered evidence and produce something a lawyer could review. The current benchmark covers more than 1,200 tasks across 24 practice areas. Individual runs can exceed 1,000 turns and consume hundreds of thousands of tokens.
Harvey did not simply teach Kimi K3 more legal facts. It taught the model a process for finishing legal work.
Teaching the model what “finished” means
LAB’s defining metric is “all-pass.” A task succeeds only when every rubric criterion passes. A diligence memo that catches nine material risks and misses the tenth is not counted as 90% complete; the whole task fails.
That standard makes the scores look low. It also makes small absolute changes consequential.
Harvey reports that Tenet reached a 19.7% all-pass rate on LAB. Its prose describes the improvement over base Kimi K3 as nine percentage points and “almost twice” as many completed tasks. On LAB Contracts, Tenet reached 11.3%, an increase of about two percentage points or roughly 22% relative to the base. The widely repeated 82% and 22% figures are relative improvements—not final accuracy rates and not percentage-point gains.
Those distinctions are not pedantry. They determine whether readers see Tenet as a meaningful improvement on a hard problem or mistakenly believe legal agents are already reliable enough to run unsupervised.
| Evaluation | Tenet result | Reported comparison | What to remember |
|---|---|---|---|
| LAB all-pass | 19.7% | About +9 percentage points versus Kimi K3; roughly an 82% relative lift in Harvey’s launch framing | Tenet used the public harness with an added finish tool; several comparator scores came from Vals |
| LAB Contracts all-pass | 11.3% | About +2 points; roughly +22% relative | A 50-task holdout, evaluated internally by Harvey; no public leaderboard at publication |
| APEX Agents, corporate law | 74.0% | Above Kimi K3 and the named frontier comparators in Harvey’s table | Harvey used direct filesystem access rather than Mercor’s canonical MCP harness; it reports Pass@1 over eight trajectories |
| RedlineBench | 55.5% | Above the Kimi K3 base, just below the strongest named comparison in Harvey’s table | Standard Crosby configuration, but some comparison scores were provided privately |
| PRBench hard subset | No statistically significant gain | Harvey reports 36.0% to 36.8% criteria pass on this separate subset | A useful null result that limits claims of universal improvement |
The generalization evidence is important. Harvey says Tenet had not seen APEX Agents or RedlineBench during training, yet improved on both. That is stronger evidence than winning only on a benchmark whose structure shaped the training environments.
But the comparisons are not laboratory-pure. Harvey’s appendix openly documents differences in harnesses, judges and score provenance. On APEX Agents, for example, its direct-bash setup improved Kimi K3 from Mercor’s published 58.8% to 67.5%. The same harness helped some comparison models and hurt others. LAB Contracts has no public leaderboard, and every reported score there came from Harvey. RedlineBench includes some privately supplied comparator results.
This is better disclosure than most launch posts provide. It still means “state of the art” should always be followed by the benchmark, configuration, evaluator and date.
Inside the Tenet training run
Harvey and Fireworks post-trained Kimi K3 with asynchronous reinforcement learning. In plain English, many agents attempted legal assignments in parallel. Their work was graded against expert rubrics, and the resulting reward signals adjusted the model toward behaviours associated with better, more complete and more efficient work.
The optimisation method was group-sequence policy optimisation, or GSPO. For each task, the system sampled a group of independent attempts, scored them, then learned from their relative performance. Near-tied groups were judged again to reduce noise. Rewards combined granular rubric coverage, a holistic count of legal issues solved and a bonus for perfect all-pass results. A smaller length term favoured concise deliverables when quality was equivalent.
Harvey used rank-64 LoRA adapters across the model’s attention, feed-forward and routed-expert weights. That is not the same as updating every parameter directly, but it is far broader than attaching a tiny adapter to one corner of the network. The research post says the setup touched roughly 500,000 expert tensors across K3’s enormous mixture-of-experts architecture.
The training dataset contained approximately 1,750 agentic environments. Each optimisation step consumed eight groups of eight rollouts; Harvey reports 150 steps and more than 10,000 individual rollouts per epoch. The run used roughly 150 NVIDIA B300 GPUs over two months.
The infrastructure problem was not only scale. Reinforcement learning assumes that probabilities computed during training correspond closely to those produced by the serving system that generated the rollouts. With a huge, sparsely routed model running at finite precision, tiny numerical differences can destabilise learning. Fireworks says its platform aligned the trainer and serving deployments at the kernel level, hot-loaded new weights without restarting the fleet, and used router replay to keep the two sides consistent. Earlier Harvey–Fireworks experiments on Kimi K2.6 had already shown that post-training and inspectable agent harnesses could improve legal-task completion at materially lower inference cost than a closed frontier model.
Harvey also says no customer data entered the Tenet post-training corpus. That is a meaningful privacy boundary. It does not by itself answer every production-security question, but it separates the research result from a model trained on confidential client matters.
Tenet is one model; Harvey’s research program is a portfolio
Harvey’s announcement describes three additional specialist capabilities that could be deployed as tools or subagents. They are easy to misreport as three Tenet variants. They are not.
| Capability | Actual base | Partner and approach | Reported result | Status |
|---|---|---|---|---|
| M&A Diligence | GLM-5.2 | Baseten; a recursive-language-model harness that lets a root agent search and delegate over huge virtual data rooms | Criteria pass rate rose from 46.1% with the harness alone to 60.1% after self-distillation training | Early research; fuller technical report promised |
| Review Table | GLM-5.2 | Applied Compute; post-training inside Harvey’s document-review environment with rewards for correctness, schema adherence, abstention and precise citations | +3.6 points in answer quality and +12.1 in citation quality at roughly one-tenth the cost per cell versus the strongest named baselines | Research tied to a production-shaped environment |
| Firm Knowledge | Qwen3.8-27B | Engram; study, structured memory, distillation and RL over a synthetic law firm | More than +15% criteria pass, nearly +10% task completion, 58% fewer tokens on completed trajectories and about 90% lower cost per query | Initial research results in a synthetic environment |
The M&A diligence work addresses tasks that can require searching up to 80 million tokens of documents. The Review Table project teaches a model when to abstain and how to cite evidence without citation spam. Engram’s Firm Knowledge research shifts computation from repeated inference into a study phase that builds durable memory of a synthetic firm.
Together they reveal Harvey’s likely end state: not one universal legal model, but a portfolio in which a general legal agent routes work to specialists trained inside the right environment.
That architecture also explains why access to open weights matters. A company can select the best available base for each job, train inside its own harness, control the serving stack and replace components without rebuilding the product around a single vendor.
The economics are about finished work, not cheap tokens
Open models are often marketed with an incomplete cost comparison: price per million tokens. Tenet’s training strategy points to the more useful equation.
Cost per successful task = token price × tokens consumed × attempts required ÷ completion rate.
A model with a lower API price can still be expensive if it searches blindly, produces bloated reasoning traces or repeatedly misses one rubric criterion. A more expensive model can be economical if it succeeds on the first attempt. Harvey trained Tenet to prefer efficient tool use and shorter trajectories when quality was equal, so it could raise completion while holding cost roughly steady.
Harvey’s quality–cost chart places Tenet on a markedly better frontier than Kimi K3 and the named closed models in its LAB setup. Yet the preview does not disclose total training cost, full deployment cost, latency or Harvey’s customer pricing. “Less than one-quarter the cost” should therefore be read as a benchmarked inference comparison under specified model configurations—not a complete total-cost-of-ownership result.
For a deeper baseline on K3’s public API pricing, architecture and independent benchmark status, see Kingy.ai’s Kimi K3 benchmark and pricing analysis.
Where the moat moved
If Harvey’s starting weights are available to competitors, what prevents another legal-AI company from copying the result?
Not secrecy around the foundation model. The moat moves into the system surrounding it:
- realistic task environments that reflect how legal work is assigned and reviewed;
- expert rubrics that turn partner judgement into trainable signals;
- reliable synthetic-data generation and human remediation;
- infrastructure capable of training and serving a multi-trillion-parameter sparse model;
- agent tools, memory and routing designed around the work;
- benchmarks that reveal which changes actually improve finished deliverables;
- distribution, trust, security and integration inside law-firm workflows.
Those assets compound. Every improved environment can produce better training traces. Every expert rubric can sharpen evaluation. Every production-shaped harness can reduce the gap between a research checkpoint and a useful agent.
The foundation model still matters enormously. But it may no longer capture all—or even most—of the application-layer value.
The open-weight policy paradox
Tenet arrives in the middle of a policy argument, but the current U.S. position is more supportive of open weights than some commentary suggests.
The White House’s 2025 AI Action Plan explicitly calls for encouraging open-source and open-weight AI. It argues that downloadable models let startups innovate without dependence on a closed provider, help organisations keep sensitive data away from third-party vendors and have geostrategic value. It also directs NTIA to encourage adoption by small and medium-sized businesses.
Tenet is almost a purpose-built exhibit for that argument—except the foundational model is Chinese.
That creates the paradox. Restricting American access to powerful foreign open weights could deny U.S. startups raw material without stopping the foreign laboratory that created it. Once a checkpoint is globally downloadable, it is also difficult to “un-release.” Moonshot can continue training in China; an American vertical company facing a domestic use restriction loses the option to adapt what already exists.
The argument should not be pushed past the evidence. Controls on advanced chips, training clusters or particular national-security deployments operate differently from restrictions on downloading weights. They may affect a foreign lab’s future capability. Procurement rules can address government risk without banning private research. Requirements for evaluations, provenance or security are not the same as a prohibition.
There are serious counterarguments. Downloadable weights can have safety layers removed. They may expand cyber, biological or fraud capabilities. A U.S. company that builds on a Chinese model inherits licensing, supply-chain and geopolitical exposure. Legal AI adds confidentiality, privilege, professional-responsibility and jurisdiction-specific accuracy requirements that no benchmark eliminates.
The Commerce Department’s NTIA open-model report offers the right discipline. It found meaningful benefits for competition, research and privacy alongside safety and national-security risks, but concluded that the evidence available in 2024 did not justify a blanket restriction. It recommended monitoring capability thresholds and assessing the marginal risk of open weights compared with closed models and existing technologies.
Tenet adds new evidence on the benefit side of that ledger. It does not settle the risk side.
The credible policy conclusion is narrower and stronger: regulation should target demonstrated capabilities, deployment contexts and downstream harms precisely. A rule that treats every downloadable model as uniquely dangerous may impose its largest immediate cost on the companies trying to turn general intelligence into accountable, domain-specific products.
What we still do not know
Harvey calls Tenet a research preview. Its technical post does not provide a public Tenet checkpoint, a standalone API, the complete training corpus or the held-out legal matters required for independent reproduction. Harvey II presents Tenet as the platform’s new legal-specific intelligence, but does not explain whether customers can select the model directly or which workflows already use it.
We also do not have published evidence on hallucination rates in live client work, jurisdiction-by-jurisdiction reliability, latency, total training expenditure or the cost of lawyer review. Benchmark completion is not permission to practise law autonomously.
Those gaps do not erase the achievement. They define the next test.
If Tenet’s gains survive independent evaluation, appear in production outcomes and extend across jurisdictions, Harvey will have demonstrated a repeatable industrial model: take frontier open weights, encode the standards of a profession, and deploy the resulting specialist at a cost that supports persistent agents.
That possibility is larger than legal technology.
The verdict
Harvey did not build Kimi K3. Moonshot did. Harvey did something that may prove just as important for the next stage of AI competition: it turned a general foundation into specialised, measurable and potentially cheaper professional capability.
The company’s strongest advantage is not possession of a secret base model. It is the ability to define legal work precisely enough that a model can practise it, fail it, receive a useful signal and improve.
Tenet therefore deserves attention for two reasons. It is an ambitious legal model with promising benchmark results. More importantly, it is evidence that American companies can build valuable frontier systems from globally available weights—provided they retain access to the models, compute, experts and freedom to experiment.
The foundation came from China. The infrastructure and legal training stack were assembled in the United States. The resulting capability belongs to neither story alone. It is a preview of an AI market in which intelligence crosses borders, while competitive advantage is rebuilt one domain at a time.
Frequently asked questions
What is Harvey Tenet?
Tenet is Harvey’s first model post-trained specifically for long-horizon legal work. It starts from Moonshot AI’s Kimi K3 base and adds reinforcement learning over legal-task environments built from public, synthetic and human-expert data.
Is Tenet open source?
Harvey calls it a post-trained open-weight model, but as of the research cutoff it had not linked to publicly downloadable Tenet weights, a Tenet-specific licence or a standalone API. Kimi K3’s base weights are public under Moonshot’s custom Kimi K3 License.
Did Tenet improve legal performance by 82 percentage points?
No. Harvey’s launch framing describes an approximately 82% relative improvement on LAB. The absolute improvement was about nine percentage points, reaching a 19.7% all-pass rate.
Are Harvey’s three specialist models also based on Kimi K3?
No. The disclosed M&A Diligence and Review Table research used GLM-5.2. Firm Knowledge used Qwen3.8-27B. Harvey describes these as separate specialist capabilities that can be deployed as tools or subagents.
Does Tenet prove open-weight models should never be regulated?
No. It provides concrete evidence that open weights can support U.S. competition and domain innovation. Safety, national-security and deployment risks still require evidence-based controls. The relevant question is which intervention addresses a demonstrated risk without unnecessarily removing useful models from legitimate downstream builders.
Editorial methodology and limitations
Kingy.ai reviewed Harvey’s Tenet research preview and appendix, Harvey’s public LAB repository and methodology posts, Moonshot AI’s Kimi K3 model card and licence, the Harvey II announcement, Fireworks’ training and K3 deployment posts, and specialist research published by Harvey, Applied Compute and Engram. Policy analysis uses the White House AI Action Plan and NTIA’s report on widely available model weights.
Benchmark values are attributed to their reporting organisation. Kingy.ai did not receive a Tenet checkpoint, run the model, inspect private training data or independently reproduce Harvey’s evaluations. Relative and absolute benchmark changes were kept separate, and product availability was limited to what Harvey publicly documented by the evidence cutoff.
Selected sources
- Harvey Tenet Research Preview and technical appendix
- Harvey Legal Agent Benchmark repository
- Introducing Harvey’s Legal Agent Benchmark
- Moonshot AI Kimi K3 model repository and report
- Kimi K3 License
- Harvey and Fireworks post-training research
- Harvey Review Table research with Applied Compute
- Engram’s Firm Knowledge research
- America’s AI Action Plan
- NTIA report on dual-use foundation models with widely available weights