Tenet shows vendor‑led post‑training + agentic RL can lift long‑horizon legal work, but the release stops short of reproducibility or open weights
Harvey announced a research preview for Harvey Tenet on August 20, 2026. Tenet is a post‑trained checkpoint built from the open‑weight Kimi K3 base, trained with Harvey’s Fireworks trainer and an asynchronous reinforcement‑learning loop targeted at long‑horizon legal agent tasks. The company reports substantial gains on its in‑house Legal Agent Benchmark (LAB), transfer improvements to some external boards, and a detailed training recipe, but it did not publish Tenet’s weights, model card, or API.
What Tenet is, in plain terms
Tenet is a post‑trained checkpoint. Harvey started from the Kimi K3 foundation model and applied targeted additional training using synthetic data, public legal corpora, and human expert traces. The training stack combined Harvey’s Fireworks rollout system with an asynchronous RL loop where rollouts and the trainer run at the same time and updated weights hot‑reload into deployments.
Harvey reports these key engineering and training choices (company disclosures):
- Reward = fraction of rubric criteria satisfied + count of legal issues solved + an all‑pass bonus for completing every atomic rubric item.
- LLM‑as‑a‑judge: Harvey reports choosing Kimi 2.6 as the judge after internal ablation studies (no public ablation report provided).
- Policy optimizer: GSPO (Group‑sequence policy optimization) with a rank‑64 LoRA applied over the full K3 network.
- Scale and scope: Harvey reports roughly 150 NVIDIA GPUs (concurrent) running for two months. About 1, 750 distinct sandboxed environments and more than 10, 000 total rollouts per epoch were used (Harvey reported). The company does not publish GPU‑hours, FLOP counts, or whether rollouts are per‑environment or aggregate.
Headline results (Harvey‑reported)
- Almost 2× held‑out task completion on LAB versus the base K3 model (Harvey‑reported).
- +20% task completion on LAB: Contracts versus base K3 (Harvey‑reported).
- LAB all‑pass rate lifted by +9 percentage points; LAB: Contracts all‑pass rate lifted by +2 percentage points (Harvey‑reported).
- LAB: Diligence pass rate reported moving from 43.8% → 60.1% after post‑training (Harvey‑reported).
- PRBench hard subset: 36.0% → 36.8% (Harvey called this not statistically significant).
- Firm‑knowledge experiment using Engram condensation: criteria pass rate +15%, task completion ≈ +10%, tokens in completed trajectories −58%, and cost per query ≈ −90% (Harvey’s proprietary “intelligence‑per‑token” metric: 190.8 vs 129.3 for their “best frontier configuration”).
- Harvey reports untrained transfer improvements to Mercor’s APEX Agents (APEX v1) and Crosby’s Redline Bench (Harvey‑reported).
“Harvey Tenet … is a research preview announced on August 20, 2026.”, Marktechpost
What an independent auditor found
Marktechpost audited Harvey’s preview and verified some numbers while flagging interpretive and comparability concerns. Their summary: 19 claims examined, 1 verified, 10 self‑reported, 6 flagged, and 2 unverifiable. They computed an “inflation score” of 78 using their published formula (Marktechpost Reality Check).
- Harvey’s evaluation harness lifts raw K3 scores: Harvey disclosed that the harness alone moves bare K3 on Mercor APEX from 58.8% → 67.5%, showing that tooling and evaluation choices can explain a material portion of reported gains.
- Public messaging framing: Harvey’s X thread presented the LAB all‑pass lift as “+82%” by changing the denominator. Marktechpost highlights that percentage framing can mislead when readers expect percentage‑point changes.
- Several claims are tied to vendor tooling, proprietary metrics, or underspecified baselines (for example, cost‑comparison claims and the “intelligence‑per‑token” metric) and therefore are not independently verifiable from the preview alone.
“Nothing Harvey published was contradicted.”, Marktechpost Reality Check lead
Product and governance realities
Tenet is a vendor‑owned research preview. Harvey published the training recipe and technical write‑ups, but did not release Tenet’s weights, a model card, or a public API. The company says the work will move “from research to production” inside Harvey’s platform over time, but it did not provide a product roadmap, timing, or on‑prem licensing commitments.
Harvey states that no customer data was used for post‑training. The preview omits full reproducibility artifacts: exact dataset lists, complete hyperparameters, and precise GPU‑hour or FLOP accounting are not published.
Harvey social posts included claims such as “Our first post-trained open-weight model” and pricing phrasing like “Less than a fourth the cost of leading foundation models.” These claims appear in company materials and social posts; Marktechpost notes some of the framing is underspecified or dependent on unnamed baselines.
Why this matters for legal teams
Long‑horizon legal tasks, M&A diligence, contract review/redlining, and multi‑document reasoning require planning, stateful memory, tool orchestration, and strict grading. Tenet’s recipe targets those constraints with long rollouts, rubriced rewards, LLM‑as‑judge grading, and specialist sub‑agents such as a Review Table and an RLM (Recursive Language Model) harness for multistep diligence. Harvey reports rollouts that sometimes exceed 1, 000 turns for certain tasks.
Three procurement‑level caveats for legal ops and AI leaders:
- Benchmarks are fragile. Small changes to the harness, judge model, or rubric can swing outcomes. Example: Harvey’s harness alone boosts K3’s APEX score materially, and percentage framing on social media converted a +9 percentage‑point improvement into “+82%”.
- Weights and governance are unresolved. If you require on‑prem hosting, audited model cards, or independent reproducibility, a closed research preview is not sufficient for procurement or compliance needs.
- Safety, privacy, and operational controls are under‑documented. The preview does not detail production safeguards for matter state, encryption, scrubbed logs, or auditability when agents operate on sensitive client files.
Technical signals engineers should notice
- Asynchronous RL + hot‑reload. Harvey’s approach runs rollouts and the trainer at the same time and injects updated weights into deployment. That reduces iteration lag but requires numerical alignment in routing and tokenization.
- MoE alignment techniques. Fireworks uses “token‑in‑token‑out” and router replay to keep MoE routing numerically stable between training and inference. That terminology addresses a real engineering risk in large Mixture‑of‑Experts systems.
- Judge circularity risk. Grading with an LLM judge (Harvey reports Kimi 2.6 after ablations) improves automated scoring but introduces circularity. The judge choice and rubric materially affect reported performance and must be disclosed for reproducibility.
- Hybrid adaptation strategy. Using a rank‑64 LoRA with GSPO over the full network is a compromise. Low‑rank adaptations limit retraining costs while optimizing policy behavior at the sequence and rollout level rather than full‑model fine‑tuning.
What buyers should do, a short checklist
- Ask the vendor for a reproducibility runbook: publish model card, weights (or audited equivalence), raw benchmark tasks and seeds, judge model and rubric, and GPU‑hour/FLOP accounting.
- Request operational safeguards: encryption and retention policies for matter memory, controls for training data leakage, and SLAs or audit access for on‑prem or air‑gapped deployments if required.
- Validate claims in your harness: run blind A/B comparisons using your evaluation harness and representative matters to separate model improvements from tooling or harness effects.
Key questions and short answers
- What is Harvey Tenet and when was it announced?
Harvey Tenet is a post‑trained checkpoint based on the Kimi K3 base, trained with Fireworks and asynchronous reinforcement learning; it was announced as a research preview on August 20, 2026 (Harvey disclosure; Marktechpost summary).
- Does Tenet outperform Kimi K3 on legal tasks?
According to Harvey’s published numbers, Tenet substantially improves task completion (Harvey reports almost 2× held‑out task completion on LAB vs base K3), +20% on LAB: Contracts, and LAB all‑pass up +9 percentage points. These are vendor‑reported results and should be validated in your environment.
- Are the model weights and API available?
No. Harvey published the training recipe and technical details but did not release Tenet’s weights, model card, or a public API; the checkpoint is a vendor‑owned research preview intended for future productization inside Harvey’s platform (Harvey disclosure).
- Did an independent auditor check the claims?
Marktechpost audited the preview, verified some figures, and flagged interpretive issues: many claims are self‑reported, some gains may reflect harness effects, and certain cost or metric claims are not independently verifiable from the preview alone (Marktechpost Reality Check).
- Will this change how law firms buy AI?
Tenet signals that targeted post‑training and agentic RL can materially improve long‑horizon legal performance. Expect increased buyer focus on evaluation harnesses, judge transparency, and operational governance; firms that require open weights or audited reproducibility should wait for published artifacts or product releases that include governance details.
Final read
Tenet is a useful field note: a vendor demonstration that post‑training a foundation model with agentic RL and a production harness can move the needle on long‑horizon legal tasks. For purchasers, it is a directional signal, not a procurement cert. Treat vendor previews as hypotheses to test. Demand reproducibility artifacts, push for judge and rubric transparency, and validate claims in your own harness before changing risk profiles or workflows.