wbstorm

Harvey Tenet explained: 2026 legal open-weight post-training results

Harvey posted Tenet, its first post-trained open-weight model. The company says it beats the Kimi K3 base on its own legal-agent bench. This is a research preview; the scores are Harvey’s.

On 20 August 2026 the legal-AI firm Harvey posted a research preview. It was not a new general model drop. It was the company’s first post-trained open-weight checkpoint aimed at long-horizon legal work. The name is Tenet.

The base is Moonshot’s Kimi K3. Post-training was done with Fireworks. The page is titled Tenet Research Preview. Law.com wrote the same day. The scores below are Harvey’s, not a third-party rerun.

Kimi K3 baseAsync RLSelf-reported LAB

Which scores the post prints

≈2×
LAB hold-out completions vs K3 (Harvey)
+20%
LAB Contracts vs K3
2 mo
Wall time on ~150 B300s
ItemHarvey’s wordingHow to read it
LAB hold-outCompletions almost doubledInternal Harvey runs
LAB Contracts~+20%, all-pass +2 ppNo public board yet
RankFirst on Contracts, second on LABCompany-reported
Customer dataNot used in post-trainingStated in the appendix
Versus the base
After synthetic, public legal, and expert data, hold-out LAB completions almost doubled; LAB Contracts rose about 20%; all-pass rates rose 9 and 2 points.
Transfer
Harvey says Mercor APEX Agents (corporate law) and Crosby’s Redline Bench, unseen in training, also beat K3. Knowledge suites such as LegalBench, CUAD, and MAUD stayed close to the base.
Cost
Reward shaping pushed shorter tool use. Harvey says quality rose while cost stayed roughly flat. Open-weight token prices are lower. No dollar rate is given.
Judge
Rollouts were graded by an LLM judge. Ablations picked Kimi 2.6.

How post-training was set

  1. 1

    Environments like LAB

    Each item is a closed matter: a short partner-style ask, client files, and an expert rubric. About 50 binary criteria on average; a rollout can pass 1,000 turns.

  2. 2

    Write to disk

    The agent searches, reads, and drafts in a sandbox, then writes deliverables to disk. Reward mixes rubric coverage, a holistic term, and a perfect-score bonus.

  3. 3

    GSPO + LoRA

    A rank-64 LoRA over the full Kimi K3 net. Eight groups of eight rollouts per step. About 1,750 environments; ~150 steps and more than 10,000 rollouts per epoch.

  4. 4

    Async loop

    Fireworks aligned train and inference kernels. About 150 NVIDIA B300 GPUs for about two months.

What is still unpublished

01

Research preview

The post says it will move work from research into the product and scale compute. It gives no public weight-download date and no external API price.

02

Harness drift

LAB used the official hold-out plus a finish tool from training. APEX Agents used an internal harness that is not Mercor’s default.

03

No Contracts board

LAB Contracts has 500 tasks; the reported split is a 50-task hold-out run by Harvey.

Limits

No independent lab is cited as reproducing Tenet on one shared harness. No hosting URL for weights. Next steps are more LAB jurisdictions and a research-to-product move — without a date.

# Sources
# Harvey, 2026-08-20: Tenet Research Preview / Update on Harvey's Post-Training Effort
# Law.com, 2026-08-20: Harvey Introduces Tenet
Can Tenet be downloaded now?

The post frames a research preview and an open-weight direction. It does not give a download page or date. Do not treat the weights as public.

Is this a new pretrain?

No. Harvey says it post-trained Kimi K3. Other capability tracks used GLM-5.2 or Qwen, not the same checkpoint.

Are the ranks a public league?

Not as such. LAB is Harvey’s open bench, but the Contracts rank and some comparisons are in-house. The APEX harness also differs from Mercor’s.

Create room