Harvey Tenet

Harvey Tenet is a legal-domain large language model derived from Kimi K3 and post-trained by Harvey with Fireworks AI. Harvey described the system in an Aug. 20, 2026 technical preview as its first post-trained open-weight model. The stated goal was to improve long-horizon legal work in which an agent must examine matter files, use tools, and produce reviewable documents.^[1]

Tenet remained a research-preview system at the Aug. 21 cutoff for this article. Harvey did not provide a public checkpoint, model card, license, repository, inference endpoint, or download instructions. Contemporaneous independent reporting said the model was not yet live in Harvey and that the company had not given a deployment date.^[1][4] Harvey's benchmark claims are therefore evidence about its disclosed research runs, not proof of public availability or performance in live legal matters.

Key facts

Scroll sideways for more →

Status and scope

Harvey introduced Tenet in two stages. Its Aug. 18 Harvey II announcement presented the model as the legal-specific intelligence behind a broader product update. The Aug. 20 research preview supplied the training and evaluation details.^[1][3] Bloomberg Law reported the model as part of a wider effort by legal-technology companies to reduce their reliance on model suppliers such as OpenAI and Anthropic.^[5]

The product language did not establish that customers could select or use Tenet. Business Insider reported on Aug. 18 that Tenet was not live and that Harvey declined to say when it would be deployed.^[4] The technical preview likewise described bringing the research into production as future work. The most supportable status is therefore a named model under development, not a generally available Harvey feature.

Harvey called Tenet open-weight. That term normally describes access to trained parameters, but no Tenet parameters or license accompanied the preview. The label records Harvey's intended release category; it does not make the model downloadable. Early independent coverage also used the term proprietary, which reflects the absence of public access at announcement rather than a published license determination.^[1][4]

Model basis

Tenet starts from Kimi K3, an open-weight mixture-of-experts model developed by Moonshot AI. Moonshot's public model summary lists 2.8 trillion total parameters, 104 billion activated parameters, 896 routed experts with 16 selected for each token, native image input, and a context length of 1,048,576 tokens.^[13]

Those specifications describe the base. Harvey did not publish a Tenet model card stating whether its serving configuration retained every Kimi K3 modality, context setting, quantization choice, or tool interface. Kimi K3's license also does not determine the license of an unpublished derivative checkpoint. Tenet should not be assigned Kimi K3's license unless Harvey releases terms for the adapted weights.

The choice of a sparse base made post-training an infrastructure problem as well as a data problem. Fireworks had previously announced private-preview support for Kimi K3 LoRA training and multi-adapter serving.^[16] That platform announcement provides context for the collaboration, but its general Kimi K3 claims do not document the Tenet run.

Training method

Environments and rewards

Harvey trained Tenet on about 1,750 agentic legal environments modeled on its Legal Agent Benchmark (LAB). Each environment contained a short instruction framed as a request from a partner, a closed set of client-matter files, and an expert rubric listing the required parts of the work product. A rollout placed the AI agent in a sandboxed workspace with tools for searching the matter, reading documents, and drafting files. The episode ended when the agent wrote its final deliverables.^[1]

LAB's public release contains task data and an execution harness. Harvey describes the initial benchmark as more than 1,200 tasks across 24 practice areas, assessed with more than 75,000 expert-written rubric criteria. A task passes under its strict all-pass rule only when every criterion passes. Tenet's reported LAB evaluation used an official holdout rather than the public task set.^[6][7]

Harvey said the training corpus combined public legal material, synthetic data, and data produced or reviewed by human experts. Mercor and other partners helped create and remediate expert datasets. Harvey also said no customer data was used in its post-training work.^[1] The preview does not publish the training corpus, task files, expert instructions, or provenance ledger, so the no-customer-data statement is a company disclosure rather than an independently audited finding.

Each rollout was scored by a rubric-based LLM-as-a-judge. The reward combined the fraction of individual criteria satisfied, a holistic count of underlying legal issues solved, and a bonus when all criteria passed. Harvey said it compared candidate judges with heavier frontier models and selected Kimi 2.6 as the best quality-efficiency tradeoff for training.^[1]

Model judging permits many trajectories to be evaluated during reinforcement learning, but it can introduce systematic error. Research by Ye and colleagues identified 12 potential LLM-judge biases and found significant bias in some tested settings.^[14] That study did not test Harvey's judge and does not show that a particular Tenet score was biased. It does mean that judge choice, prompts, agreement with lawyers, and repeated grading are part of the evidence needed to interpret the results.

Optimization and compute

Harvey used Group Sequence Policy Optimization (GSPO). The published GSPO method computes importance ratios and clipping at the sequence level, and was designed in part to improve the stability of reinforcement learning for sparse models.^[12] For Tenet, each training task produced a group of independent rollouts. The system normalized advantages within the group, rejudged near ties, and used a small length term to prefer shorter deliverables when their substantive reward was similar. Harvey also reported sequence-averaged importance weights, double-sided clipping, and removal of tokens with unusually high importance ratios.^[1]

The model update used rank-64 LoRA across the full Kimi K3 network rather than adapting only a small set of output layers. Harvey described the coverage as all attention, MLP, and routed-expert weights, or about 500,000 expert tensors. AdamW was the optimizer. Each optimizer step consumed eight task groups with eight rollouts per group.^[1]

Scroll sideways for more →

Rollout generation and optimization ran asynchronously. A local orchestrator kept work in flight while the trainer consumed completed groups, with a staleness limit intended to bound how far samples could drift from the current policy. Updated weights were hot-loaded into the serving deployments after each optimizer step. Harvey said Fireworks co-designed training and rollout kernels to reduce numerical differences between the two systems and used router replay to keep sparse-expert routing aligned.^[1]

These disclosures identify the broad recipe and scale, but they are not a complete reproduction package. The preview does not provide the checkpoint, environment corpus, exact prompts, full hyperparameter set, generated trajectories, training logs, or a detailed compute and cost ledger.

Evaluation

LAB and LAB Contracts

Harvey's results chart reports a 19.7 percent all-pass rate for Tenet on LAB and 11.3 percent on LAB Contracts. The technical preview gives the corresponding absolute changes from the Kimi K3 base as 9 and 2 percentage points. Harvey's social announcement expresses the same comparisons as relative increases of 82 percent and 22 percent.^[1][2]

Scroll sideways for more →

The two ways of stating each gain are not interchangeable. A 9 percentage-point increase from a low base can be an 82 percent relative increase. It is not an 82 percentage-point increase. The same distinction applies to the 2 percentage-point and 22 percent figures.

Harvey ran Tenet in the public LAB harness with an added finish tool that had also been used during training. The company used baseline scores from Vals because they tracked its internal method more closely, while acknowledging that its harness and judge differed from other evaluators.^[1] Vals' Aug. 19 snapshot used its own infrastructure and two judges, and reported Kimi K3 at 10.83 percent task resolution and 90.79 percent criteria pass.^[8] Those Vals values illustrate how all-pass can remain low even when most criteria pass. They should not be substituted for Harvey's same-harness comparison.

LAB Contracts extends the format to drafting, review, and negotiation. The public benchmark has 500 tasks; Harvey used an official 50-task holdout. Because there was no public LAB Contracts leaderboard, every displayed score in the Tenet preview was an internal Harvey run.^[1]

Other legal benchmarks

The preview also reports evaluations outside LAB, but the execution method varies by benchmark.

On the corporate-law portion of APEX Agents, Harvey reported 74.0 percent for Tenet. Mercor's public canonical leaderboard lists Kimi K3 at 58.8 percent.^[9] Harvey instead mounted the task filesystem directly and gave the model Bash access. In that harness, Kimi K3 rose to 67.5 percent. The change helped some comparator models and hurt others. Harvey retained Mercor's judge and Pass@1 over eight trajectories, but it used published canonical scores for the comparison table. The 74.0 result is therefore informative about Tenet's run, not a uniform same-harness ranking of every listed model.^[1][9]

Mercor separately ran Tenet on the held-out 100-task legal subset of APEX v1. Harvey's chart reports 73.0. Mercor did not disclose the tasks, runs, or task-level scores to Harvey. This gives the result a benchmark-owner execution boundary, but it is not a public replication because no Tenet checkpoint is available.^[1]

For Crosby RedlineBench, a multi-turn contract-negotiation benchmark with attorney-authored rubrics, Harvey said it used Crosby's standard Harbor configuration and judging setup. The chart reports 55.5 for Tenet.^[1][10] For Scale's PRBench, Harvey used the standard configuration. On the 250-question hard subset, the criteria-pass rate changed from 36.0 for Kimi K3 to 36.8 for Tenet, and Harvey explicitly reported that difference as not statistically significant.^[1][11] The external suite therefore does not support a claim that post-training improved every benchmark.

Harvey also tested legal knowledge rather than agent execution. Its LegalBench run omitted the hand-graded Rule QA task and used structured outputs for exact-match items. For CUAD, the company reported generative F1 instead of the dataset's log-probability-based headline metric. For MAUD, it sampled up to 20 examples per question and used generative answer matching rather than AUPR.^[1][15] These implementation notes matter because scores produced under different output constraints and metrics are not direct replacements for the original benchmark reports.

Cost claims

Harvey said its reward design preferred trajectories that used fewer tokens when substantive performance was equal. Its social post described Tenet as operating at less than one-quarter the cost of leading foundation models, and the technical preview plotted a quality-cost frontier.^[1][2]

No public Tenet price, serving endpoint, full token ledger, hardware-utilization record, or reproducible deployment configuration accompanied that claim. Inference cost depends on token mix, model pricing, batching, hardware, quantization, latency targets, and the agent harness. The less-than-one-quarter figure should be read as Harvey's calculation for its tested configurations, not a general price guarantee.

Availability and deployment

Harvey's technical page did not link public weights or a license. At the Aug. 21 research cutoff, a Hugging Face public-model query for the exact name returned no models, and a GitHub public-repository query returned no matching repositories.^[17][18] These are dated inventory checks. They do not show whether Harvey will publish an artifact later.

The absence of a checkpoint also prevents independent reproduction of the reported training and evaluation runs. It is not possible to verify the adapter contents, confirm the model configuration, run alternative judges, or test the claimed cost profile from the published materials alone. Calling Tenet open-weight describes Harvey's announcement; it should not be paraphrased as already released or open source.

Harvey II also introduced matter context, memory, permissions, and other product features.^[3] Those features are part of the surrounding product, not demonstrated properties of the Tenet weights. Similarly, a model benchmark does not verify the confidentiality, access-control, audit, or data-retention behavior of the system in which the model might later run.

Separate specialist research

The Tenet preview describes three additional capabilities for M&A diligence, Review Table, and firm knowledge. Harvey says these were trained separately and could be deployed as tools or subagents. They are not disclosed as modules inside one Tenet checkpoint.^[1]

The diligence experiment used GLM-5.2 in a recursive-language-model harness developed with Baseten. The Review Table work used a separately post-trained GLM-5.2 with Applied Compute. The firm-knowledge study used Qwen3.8-27B with Engram. Because the base models and training partners differ, their results should not be reported as Tenet benchmark scores or evidence about Tenet's architecture.

Limitations and legal-use boundary

Most Tenet results were generated or assembled by Harvey, the system's developer and the creator of LAB. The company disclosed several harness changes and metric substitutions, which helps interpretation, but it did not publish raw outputs, per-task scores, judge prompts, attorney-judge agreement, complete confidence intervals, or a preregistered evaluation plan. The PRBench hard-subset note is the one explicit statistical caveat in the preview.^[1]

No independent group had reproduced the full evaluation suite by the cutoff. Independent reporting confirmed the model announcement and its not-yet-live status; it did not validate the scores.^[4][5] Mercor's held-out APEX v1 run is a stronger separation of benchmark execution, but the lack of weights and task-level records still prevents public replication.

The evaluations use simulated or closed-universe matters and rubric-defined deliverables. They do not establish jurisdiction-specific legal accuracy, reliability on live client files, resistance to fabricated citations, confidentiality compliance, or autonomous competence. Any later operational use would still require qualified lawyer review, source verification, matter-specific permissions, and compliance with professional duties.

References

1. ^Calvin Qi, Vasudha Rengarajan, Julio Pereyra, Niko Grupen, and Gabe Pereyra. "Harvey Tenet Research Preview." Harvey, Aug. 20, 2026. harvey.ai/...post-training-update-harvey-tenet
2. ^Harvey. "Introducing Tenet, our first model post-trained for legal." X, Aug. 20, 2026. x.com/...2090454750059958440
3. ^Harvey Team. "Introducing Harvey II." Harvey, Aug. 18, 2026. harvey.ai/...introducing-harvey-ii
4. ^Melia Robinson. "Harvey's First LLM for Legal Work Is Here." Business Insider, licensed Taiwan edition, Aug. 18, 2026. businessinsider.tw/...5982
5. ^Daniel Seiden. "Legal Tech AI Firms Shift Away From Anthropic, OpenAI Reliance." Bloomberg Law, Aug. 19, 2026. news.bloomberglaw.com/...anthropic-openai-reliance
6. ^Harvey AI. "Harvey LAB: The Legal Agent Benchmark." GitHub repository, version 1.0, accessed Aug. 21, 2026. github.com/...v1.0
7. ^Niko Grupen, Gabe Pereyra, and Julio Pereyra. "Open-Sourcing Harvey's Long Horizon Legal Agent Benchmark." Harvey, May 6, 2026. harvey.ai/...oducing-harveys-legal-agent-benchmark
8. ^Vals AI. "Harvey's Legal Agent Benchmark." Updated Aug. 19, 2026. vals.ai/...hlab
9. ^Mercor. "APEX-Agents: Corporate Lawyer." Accessed Aug. 21, 2026. mercor.com/...corporate-lawyer-agent
10. ^Crosby Intelligence. "Crosby RedlineBench." Accessed Aug. 21, 2026. intelligence.crosby.ai
11. ^Scale AI. "PRBench: Professional Reasoning Benchmark." GitHub repository, accessed Aug. 21, 2026. github.com/...PRBench
12. ^Chujie Zheng et al. "Group Sequence Policy Optimization." arXiv:2507.18071, July 24, 2025. arxiv.org/...2507.18071
13. ^Moonshot AI. "Kimi K3: Open Frontier Intelligence." GitHub repository and technical report, accessed Aug. 21, 2026. github.com/...Kimi-K3
14. ^Jiayi Ye et al. "Justice or Prejudice? Quantifying Biases in LLM-as-a-Judge." arXiv:2410.02736, Oct. 4, 2024. arxiv.org/...2410.02736
15. ^Neel Guha et al. "LegalBench: A Collaboratively Built Benchmark for Measuring Legal Reasoning in Large Language Models." arXiv:2308.11462, Aug. 21, 2023. arxiv.org/...2308.11462
16. ^Fireworks AI. "Make Kimi K3 Yours: LoRA Training on Fireworks." July 26, 2026. fireworks.ai/...K3-LoRA-Training
17. ^Hugging Face. "Public model index search for Harvey Tenet." Queried Aug. 21, 2026. huggingface.co/...models
18. ^GitHub. "Public repository search for Harvey Tenet." Queried Aug. 21, 2026. api.github.com/...repositories