Harvey推出Harvey Tenet：基于Kimi K3底座，采用Fireworks进行后训练，面向长周期法律智能体工作

Harvey Introduces Harvey Tenet: A Kimi K3 Base Post-Trained with Fireworks for Long-Horizon Legal Agent Work

Harvey has released Harvey Tenet, its first post-trained model, as a research preview as of today. Tenet is a Kimi K3 base post-trained with Fireworks through asynchronous reinforcement learning on long-horizon legal work. The training corpus combined synthetic data, publicly available legal data, and human expert data. Harvey states no customer data was used. Against the base K3 model, Tenet completes almost twice as many held-out tasks on Harveys Legal Agent Benchmark (LAB) and 20% more on LAB: Contracts, raising all-pass rate by 9 and 2 percentage points respectively. Harvey reports state-of-the-art on LAB: Contracts and second place on LAB. The gains also transferred, untrained, to Mercors APEX Agents and Crosbys Redline Bench. The stated goal is twofold: build frontier legal intelligence on open-weight models, and give law firms a path to own their own specialized models.

Not yet, Harvey Tenet is a research preview announced on August 20, 2026. Harvey has not published weights, a model card, or an API endpoint. The base model is open-weight; Tenet itself is Harveys own checkpoint, and the company says the work will move from research to production inside Harveys products over time. What ships today is the recipe, not the artifact.

Company tier: Enterprise only. Access runs through Harveys platform, which is sold to law firms, mid-sized firms, and in-house legal teams. A lab with an RL stack could reproduce the method; training used roughly 150 NVIDIA B300 GPUs over two months.

Industries: Legal services, corporate in-house legal, private equity and investment banking (M&A diligence), plus regulated sectors where contract volume drives cost — insurance, financial services, healthcare, energy.

Applications: M&A due diligence memos over datarooms, contract drafting, review and redlining, structured extraction across up to 10,000 documents, and precedent search over a firms accumulated knowledge.

Against the base K3 model, Tenet completes almost twice as many held-out tasks on Harveys Legal Agent Benchmark (LAB) and 20% more on LAB: Contracts, lifting all-pass rate by 9 and 2 percentage points respectively. Harvey reports state-of-the-art on LAB: Contracts and second place on LAB, using base-model scores from Vals.

The more interesting result is transfer. Tenet also improves substantially on Mercors APEX Agents (corporate law) and Crosbys Redline Bench — neither seen during training — while holding performance on knowledge benchmarks including LegalBench, CUAD, MAUD, and Scales PRBench. Agentic training did not erode textbook legal reasoning.

Cost is co-optimized rather than traded away. Open weights lower price per token; reward shaping that prefers shorter trajectories at equal quality lowers tokens consumed. Harvey reports significant quality gains at stable cost.

Training used asynchronous reinforcement learning in sandboxed legal environments built like LAB tasks: a partner-style instruction averaging about 50 words, a client matter of key and peripheral documents, and an expert rubric of atomic pass/fail criteria — roughly 50 per task, hundreds at the extreme. A single rollout can exceed 1,000 turns.

Rollouts are graded by LLM-as-a-judge; ablations settled on Kimi 2.6. Reward combines the fraction of rubric criteria satisfied, a holistic count of legal issues solved, and an all-pass bonus. The policy is optimized with GSPO using a rank-64 LoRA over the full K3 network, eight task groups of eight rollouts per optimizer step, across ~1,750 environments and >10,000 rollouts per epoch. Fireworks co-built trainer and rollout deployments at the kernel level, with token-in-token-out and router replay, to keep a large MoE numerically aligned across training and inference.

Three capabilities trained separately

Harvey team also post-trained specialist models that Tenet can route to as tools or sub-agents:

M&A diligence: On LAB: Diligence, a single task can traverse up to 80M tokens; no baseline passed more than 43.8% of criteria. With Baseten, Harvey moved to a Recursive Language Model harness where a root agent holds the dataroom in a REPL and delegates to sub-agents. A GLM-5.2 orchestrator alone reached 46.1%; post-training it in that harness via self-distillation reached 60.1%.

Review Table: With Applied Compute, a post-trained GLM-5.2 improved answer quality by 3.6 points and citation quality by 12.1 points at roughly one-tenth the cost per cell, learning to abstain when a question does not apply.

Firm knowledge: With Engram, a Qwen3.8-27B model studies ~100M tokens of client matters into 1M tokens of structured knowledge plus parametric memory. Criteria pass rate rose more than 15%, tokens in completed trajectories fell 58%, and cost per query dropped roughly 90% — 190.8 intelligence-per-token versus 129.3 for the best frontier configuration.

Marktechpost Independent Test Facts

Harvey Tenet benchmark claims, verified

Audited: Harvey research preview, Aug 20, 2026 and the Harvey X thread Mode: default

Nothing Harvey published was contradicted. The score is high because Tenet appears on no public leaderboard not Vals, not Artificial Analysis, not Mercor. Score formula: (8 6 flags) + (15 0 contradicted) + (3 10 self-reported) = 78.

ClaimNumberIndependent checkVerdict

Completes ~2 more LAB held-out tasks than Kimi K3 base

Same result stated as +82% on X

State-of-the-art on LAB: Contracts

Benchmark owned, run and graded by Harvey

Vals #1 is Muse Spark 1.1 at 20.00%; Harvey-run, tool delta never quantified

Kimi K3 base on APEX Agents, corporate law

58.8% (Kimi K3 Max) matches exactly

Tenet substantially beats K3 base on APEX Agents

Harvey harness alone: 58.8% 67.5%

Beats K3 base on Crosby Redline Bench

Absent from the public leaderboard

APEX v1 Big Law Associate held a knowledge benchmark, not the agentic board

Blind run commissioned from Mercor; not posted to the public v1 board

Holds on LegalBench, CUAD, MAUD

Non-canonical metrics, applied to all models

Harvey calls it not statistically significant

LAB: Diligence criteria pass rate

Firm Knowledge intelligence-per-token

Engram write-up; metric is Harveys own

Our first post-trained open-weight model

Business Insider: proprietary, in-house

Less than a fourth the cost of leading foundation models

150 NVIDIA B300 GPUs, 2 months, GSPO + rank-64 LoRA

No customer data used in post-training

F1 Denominator gameThe blog reports +9 and +2 percentage points. The X thread reports the same result as +82% and +22%. Both true; the social number sounds nine times larger.

F2 Self-report as factSOTA on LAB: Contracts is a win on Harveys own benchmark. Harvey states there is no public leaderboard for it and that all scores are internal Harvey runs. LAB launched deliberately without a leaderboard.

F3 Settings mismatchHarvey disclosed this plainly: Tenet ran in the standard public harness plus a finish tool carried over from training, while rival scores came from Vals. The flag is about comparability, not concealment Harvey never published LAB with and without the tool, so its value is unquantified. Harveys own APEX figures show a harness change moving bare K3 by 8.7 points, and the LAB claim is a rank where Vals leaders sit between 12% and 20%.

F4 Settings mismatchOn APEX Agents, Tenet ran in Harveys internal bash harness while rivals used Mercors published numbers. Harvey discloses the harness lifts bare K3 from 58.8% to 67.5% within 0.1 pt of leader Fable 5 at 67.4%, before any training.

F5 FramingOpen-weight describes the Kimi K3 base, not Tenet. No weights, model card or API were published, yet multiple outlets ran headlines calling Tenet itself an open-weight release.

F6 Denominator gameRoughly one-tenth the cost per cell is measured against unnamed strongest baselines, with no serving config, precision or hardware given for either side.

F7 Denominator gameLes