Harvey has introduced Tenet, its first model post-trained for legal work. The model starts from the open-weight Kimi K3 base and is trained to complete long-horizon legal tasks inside an agent harness, rather than answer isolated legal questions.
The company announced Tenet on 20 August 2026 through an official Harvey post on X and a detailed X Article by co-founder Gabe Pereyra. Harvey reports a large improvement over the Kimi K3 base on its own Legal Agent Benchmark, alongside lower task costs and three specialist capabilities for diligence, review tables and firm knowledge.
Those results are significant enough to merit attention, but they are first-party research results. Tenet should be read as a promising domain post-training programme—not as independent proof that a legal agent is ready to practise without supervision.
What Harvey actually announced
Tenet is not a new foundation-model architecture. It is a Kimi K3 checkpoint adapted with Harvey and Fireworks for legal agent work. Harvey says the training mix combined publicly available legal material, synthetic examples and data prepared or reviewed by human experts. The company states that no customer data was used in the post-training effort.
The training environment is closer to a matter workspace than a conventional question-and-answer benchmark. An agent receives a short partner-style instruction, a closed set of client-matter files, tools for searching and drafting, and an expert rubric tied to the required deliverable. Some tasks run for more than 1,000 turns and consume hundreds of thousands of tokens.
Harvey describes the training run as:
- roughly 1,750 agentic legal task environments;
- group-sequence policy optimisation with a rank-64 LoRA over Kimi K3;
- more than 10,000 rollouts in each training epoch;
- approximately 150 NVIDIA B300 GPUs over two months;
- an asynchronous loop that updates serving weights without restarting deployments; and
- reward shaping that favours substantive rubric coverage as well as token efficiency.
This matters because the model is being trained in the same kind of loop in which it will operate: searching files, using tools, building context and producing a reviewable work product. It is a different objective from increasing legal knowledge in the model’s parameters alone.
The headline benchmark gain needs two numbers
Harvey’s X announcement says post-training increased Tenet’s all-pass rate by 82% on LAB and 22% on LAB Contracts relative to the Kimi K3 base. The longer technical article provides the absolute change: nine percentage points on LAB and two percentage points on LAB Contracts. It also describes the result as almost twice as many held-out LAB tasks completed and about 20% more LAB Contracts tasks completed.
Both ways of stating the change can be correct, but the absolute movement is essential context. “All-pass” is also a strict metric: a task passes only if the produced work satisfies every criterion in its expert rubric. It should not be confused with ordinary token accuracy or the percentage of individual rubric items met.
| Harvey-reported evaluation | Change or result | Important qualification |
|---|---|---|
| LAB hold-out | All-pass up 9 percentage points; about 82% relative | Standard public harness plus a finish tool used during training |
| LAB Contracts hold-out | All-pass up 2 percentage points; about 22% relative | Harvey internal runs on a 50-task official hold-out set |
| Overall standing | State of the art on LAB Contracts; second on LAB | Comparative scores mix disclosed external and Harvey runs |
| PRBench hard subset | 36.0% to 36.8% criteria pass rate | Harvey says the increase was not statistically significant |
Harvey also reports transfer to Mercor’s APEX Agents corporate-law subset, Crosby’s Redline Bench and Scale’s Professional Reasoning Benchmark. The appendix is unusually helpful about methodology differences: for APEX Agents, Harvey used direct filesystem and shell access rather than Mercor’s canonical structured MCP harness. Harvey says this alternative harness improved some comparison models and reduced others, which is exactly why a leaderboard position cannot be separated from its execution setup.
Tenet is more than one generalist checkpoint
The broader research programme includes three specialist capabilities that Tenet can call as subagents. Each targets a workload where a general model tends to waste context, miss structure or become prohibitively expensive.
| Specialist capability | Research setup | Harvey-reported result |
|---|---|---|
| M&A diligence | GLM-5.2 in a recursive-language-model harness developed with Baseten | Criteria pass rate rose from 46.1% to 60.1% after self-distillation post-training |
| Review Table | GLM-5.2 post-trained with Applied Compute for structured extraction across up to 10,000 documents | Answer quality up 3.6 points and citation quality up 12.1 points at roughly one-tenth the cost per cell |
| Firm knowledge | Qwen3.8-27B trained with Engram to build parametric memory and structured notes | More than 15% higher criteria pass rate, 58% fewer tokens on completed runs and 90% lower cost per query |
These are not three variants of Tenet with identical weights. They are specialist models and harnesses that the broader agent system can route to. The announcement therefore supports a systems-level reading: the product advantage may come from choosing the right model, workspace, memory and tool loop for each stage of legal work, rather than forcing one general model to do everything.
That is consistent with the economics already visible in our Kimi K3 release brief. A very large open-weight base can be expensive to host, but targeted post-training and efficient trajectories may still make each completed task cheaper than repeated calls to a closed frontier model.
“Open-weight” does not yet mean Tenet is downloadable
Harvey describes the research as building legal intelligence with open-weight models, and Kimi K3’s full base weights are available under Moonshot’s licence. The announcement does not, however, identify a public Tenet checkpoint, model card or download repository.
The careful interpretation is that Tenet is post-trained from an open-weight base, not that Harvey has released Tenet’s adapted weights. That distinction affects reproducibility, procurement and the claim that a law firm can “own its intelligence”. A customer may gain the commercial and customisation benefits of an open-model stack without receiving a checkpoint it can independently host or audit.
Buyers should ask which layer they control:
- the underlying Kimi K3 weights;
- Harvey’s LoRA or other adapted parameters;
- the legal training data and expert rubrics;
- the agent harness, tools and retrieval policy;
- the firm-specific model or memory produced later; and
- the right to export, redeploy or delete each of those artefacts.
The wider Kimi K3 architecture guide explains why the base model is technically suited to long-context agent work. Tenet’s news is the domain-specific training layer and the evidence Harvey is assembling around it.
What the cost claim does—and does not—show
Harvey says Tenet operates at less than one-quarter of the cost of leading foundation models. Its published chart uses cache-aware cost per task, and the technical article says reward shaping encouraged trajectories that used fewer inference tokens for equivalent performance.
Task cost is a better operating measure than a provider’s list price per million tokens, especially when legal agents may search and revise for hundreds of turns. Yet it is still sensitive to caching assumptions, batch volume, hardware utilisation, retries, timeouts and the number of failed tasks that must be rerun or repaired by a lawyer.
A credible buyer test should therefore report:
- completed matters or deliverables, not only token price;
- lawyer correction time and the severity of missed issues;
- uncached and cache-aware cost under the buyer’s traffic pattern;
- latency and queue time for both ordinary and long-running matters;
- repeat-run variance on the same instruction and documents;
- abstention and escalation when evidence is incomplete; and
- total cost after review, remediation, security and platform overhead.
This follows the same principle as our guide to AI model change control: a model upgrade is operationally better only when the whole workflow improves against a frozen baseline.
Why Tenet matters for legal AI
Most legal copilots begin each task with a prompt and a bundle of documents. Harvey’s direction is more ambitious: agents operate inside a matter, inherit its tools and boundaries, and learn from work patterns that can persist over time. The separate Harvey II product announcement connects Tenet to Spaces, personal memory, matter permissions and ethical walls.
If Harvey’s findings reproduce in production, three changes follow. First, domain post-training can make a capable open base materially better at complete legal work products, not just legal trivia. Second, training the agent inside its eventual harness may improve both quality and efficiency. Third, legal teams may be able to build differentiated models around their own workflows instead of relying on the same general endpoint as every competitor.
The limiting word is if. Harvey’s full post-training report is transparent about several harness differences and one non-significant result, but independent replication, jurisdictional testing and live error evidence are still missing. LAB tasks are closed-universe simulations; real matters contain ambiguous instructions, missing documents, changing facts and professional duties that cannot be reduced to a benchmark rubric.
AIEngine verdict
Tenet is one of the more consequential legal-AI announcements of 2026 because it joins an open-weight frontier base, domain reinforcement learning, specialist subagents and a serious attempt to measure complete legal work. The result points towards legal agents that are trained as systems, not chat models with a legal prompt.
The announcement is not a licence to remove lawyers from review, and it is not yet an open release of Tenet’s adapted weights. Firms evaluating Harvey should request exact model and harness versions, reproduce results on their own matters, measure lawyer correction effort, verify data isolation, and preserve a human decision-maker for substantive legal conclusions.
For now, Tenet is best understood as credible evidence that domain post-training can move the cost-quality frontier for legal agents—and as a challenge for the rest of the market to publish equally detailed evaluation methods.
