Harvey Tenet [What It Changes]

TL;DR

• Harvey Tenet is Harvey's first in-house legal model: a post-trained copy of Moonshot AI's open-weight Kimi K3, announced as of August 20, 2026, for long-horizon lawyer work rather than generic chat.

• Harvey is building Tenet so it can stop paying OpenAI, Anthropic, and Google for every token; The Next Web reports the company already runs at more than $350mn annualised.

• Business Insider reports Tenet is not live in the product yet and Harvey will not name the firms testing it.

• A two-person shop does not buy Tenet this quarter; if a model reads privileged files, treat model choice, data retention, and review queues as owned workflow steps, not as a chatbot tab.

Key Takeaways

• Harvey Tenet is a legal post-train on a Chinese open-weight base, not a replacement for Clio, Smokeball, or MyCase.

• Owning the model turns Harvey's largest variable cost into a fixed training bill; small firms feel the same meter when they paste a lease or intake packet into a general chatbot.

• Harvey's research preview says Tenet completes almost twice as many held-out Legal Agent Benchmark tasks as base Kimi K3 and raises all-pass by 9 percentage points.

• Privilege and EU AI Act deployer duties do not disappear because the weights are "open."

• Wire any later model swap into an existing document route rather than buying a second legal stack.

What Harvey Tenet is

Harvey Tenet is Harvey's first proprietary model for legal work: Kimi K3, post-trained with Fireworks AI on synthetic matters, public legal data, and lawyer-written mock files, so agents can draft, cite, and review over a long matter instead of answering one prompt at a time.

If you run a two-truck HVAC shop, a ten-person marketing agency, or a solo clinic, you will not get a Harvey Tenet login. You should still care, because you already rent the same kind of intelligence every time a general model reads a job packet, a contract, or a patient form. Harvey decided the invoice from OpenAI, Anthropic, and Google was the risk, so it trained its own model.

A solo family practice comparing MyCase vs Clio Manage lives the miniature version of that bill every night a parenting plan is pasted into a chatbot. A transactional boutique weighing Smokeball vs Clio Manage lives it when a diligence folder hits the inbox after 6 p.m.

Harvey Tenet does not file the appearance or hold the licence. It is a model. The practice system and the human who signs the letter still sit around it.

What shipped, and what did not

Harvey published the Harvey Tenet research preview on August 20, 2026. Two days earlier it had framed Tenet as the intelligence layer inside Harvey II, the product release that also added matter-scoped agents and a Memory feature for style and preferences.

According to Business Insider, Harvey introduced Harvey Tenet on Tuesday, August 18, 2026 as its first in-house proprietary model, and Tenet is not live in Harvey yet.

A research preview is not a production routing change. Pereyra would not name the law firms that might be testing it. Treat every benchmark below as a lab result until Harvey puts the model on a named default path.

On the Harvey homepage, the company states 2,400+ legal organizations build on Harvey, with 200,000+ professionals, use in 70+ countries, and 75+ AmLaw 100 firms. Those figures describe Harvey the product, not Tenet the model.

Why Harvey built it now

Harvey spent years routing customer work through models from OpenAI, Anthropic, and Google. The Next Web notes that OpenAI is an investor in Harvey alongside Sequoia and Andreessen Horowitz, so Harvey is reducing dependence on a backer whose API it also pays.

According to The Next Web, Harvey is in talks to raise at least $500m at a $15.5bn valuation, up 40% in five months, after annualised revenue moved from $190m in January to more than $350m.

Harvey now books more than $350 million annualised. That figure sits in the same TNW valuation report as the 44-times revenue multiple the outlet applies to the $15.5bn talks.

According to CNBC, Harvey raised $200 million at an $11 billion valuation on March 25, 2026, after an $8 billion round in December, and it counted more than 100,000 lawyers across 1,300 organizations at that time.

Customer counts moved between March and the August homepage. The cost structure problem did not. Tenet is the attempt to pull a slice of the frontier-API bill in-house.

The competitive clock is also running. Business Insider reports that Anthropic is chasing lawyers with plugins and that OpenAI hired Ironclad founder Jason Boehmig to lead a legal push. Harvey's answer is to own a legal-shaped model.

Europe's rival is on a parallel raise. The Next Web reports Legora is seeking funds at more than $10bn after a March Series D at a $5.55bn valuation, and that it crossed $100m ARR around April across more than 1,000 customers in some 50 markets.

Sources: CNBC; The Next Web; Harvey.

How Tenet is trained

The mechanism is post-training, not a new pretrain from scratch. Harvey starts from Kimi K3. Lawyers on staff and contractors through Mercor and Snorkel invent mock disputes and case files, then grade how models reason through them. Harvey and Fireworks run reinforcement learning in sandboxed matters that look like partner-to-associate requests.

The Harvey Tenet research preview dated August 20, 2026 is the primary technical source. It says Tenet is a Kimi K3 base post-trained with Fireworks research for long-horizon legal work, that Harvey used no customer data in post-training, and that training ran group-sequence policy optimization with a rank-64 LoRA over the full Kimi K3 network.

Further details from that same preview: about 1,750 agentic legal task environments, more than 10,000 rollouts, and training on approximately 150 NVIDIA B300 GPUs over 2 months. A task contains 50 rubric criteria on average. A single rollout can span more than 1,000 turns.

Harvey had already published the evaluation harness. On May 6, 2026 it open-sourced the Legal Agent Benchmark: more than 1,200 agent tasks across 24 practice areas, graded by over 75,000 expert-written rubric criteria. Harvey's research index lists the Tenet preview as the August 20, 2026 model card. An earlier Fireworks LAB run on June 3, 2026 put Claude Opus 4.7 at 14 / 100 all-pass and $954 on a 100-task slice, versus 18 / 100 and $368 for a GLM 5.1 worker with Opus as a sparse advisor — a K2.6-era cost story Tenet is meant to finish.

Benchmarks Harvey is willing to print

According to the Harvey Tenet research preview, the post-trained model successfully completes almost twice as many held-out LAB tasks and 20% more on LAB contracts than base Kimi K3, increasing all-pass rate by 9 and 2 percentage points respectively.

Tenet lifts LAB all-pass by 9 percentage points. That sentence is Harvey's own hold-out claim on the research preview, not a third-party audit.

The same preview says Tenet places first on LAB Contracts and second on LAB. On Mercor's APEX Agents corporate-lawyer subset, Harvey reports that running Kimi K3 in its harness improves the published 58.8% score to 67.5%. On Scale's PRBench hard subset, it reports a move from 36.0% to 36.8%, which the authors call non-statistically significant.

Harvey still evaluates older short-horizon suites. LegalBench is 162 tasks from 40 contributors. CUAD is 13,000+ labels on 510 commercial contracts covering 41 clause types. LAB is the long-horizon test.

Business Insider notes the conflict: a vendor that trains on its own benchmark is taking an exam after helping write the answer key. Read the 9-point lift as Harvey's measurement, then wait for a firm that re-runs the hold-out.

Source: Harvey Tenet Research Preview, August 20, 2026. Diligence, Review Table, and firm-knowledge rows are adjacent post-trains in the same article, not all a single K3 checkpoint.

The Kimi K3 base, and why the licence is not free

Harvey did not pick a US open-weight model. It picked Moonshot's Kimi K3.

According to Business Insider, Kimi K3 has 2.8 trillion parameters, Moonshot planned to release weights by July 27, 2026, and the API list price is $3 per million input tokens and $15 per million output tokens.

Kimi K3 lists at $3 per million input tokens. That Business Insider price sits against the same article's figures of $5 / $30 for OpenAI's GPT-5.6 Sol and about $10 / $50 for Anthropic's Claude Fable 5.

Source: Business Insider, July 17, 2026.

Open weights are not an unbounded commercial licence. According to The Next Web, Kimi K3 requires a separate agreement with Moonshot for model-as-a-service operators above $20mn of revenue in any 12 months, and Harvey already runs at more than $350mn annualised.

Harvey is not a hobby fine-tune. It is a MaaS operator well above Moonshot's stated threshold, which means the "open" base still has a commercial counterparty in Beijing.

USTA analysis: how far Harvey sits over the Kimi MaaS line

USTA analysis, using only the two figures in that TNW sentence: Harvey's >$350mn annualised revenue divided by Moonshot's $20mn MaaS threshold equals more than 17.5 times the revenue level that triggers a separate Moonshot agreement. Inputs: $350,000,000 / $20,000,000 = 17.5. If the run-rate is strictly above $350mn, the multiple is strictly above 17.5.

Privilege, privacy, and the EU file

Harvey's security page states that customer data is hosted in Microsoft Azure, that processing can sit in the EU and Switzerland, the US, or Australia, that Harvey does not use inputs, outputs, or uploaded documents to train underlying models, and that it requires Zero Data Retention by model providers. It also states annual SOC 2 Type II and ISO 27001 audits, and lists CCPA, GDPR, ISO 27701, and ISO 42001 among its public control marks.

The California CCPA page, updated August 28, 2026, still sets for-profit thresholds at over $25 million gross annual revenue, 100,000 California residents whose personal information is bought, sold, or shared, or 50% of revenue from selling that information, with statutory damages up to $750 per incident after a qualifying breach. A small US firm using legal AI is usually the customer, not the $25 million business.

On the EU side, the Commission's AI Act overview states that Regulation (EU) 2024/1689 entered into force on 1 August 2024 and became applicable on 2 August 2026, with GPAI rules effective in August 2025, transparency rules in August 2026, and high-risk obligations from 2 December 2027. Administration of justice sits on the high-risk list. TNW argues a post-train is likely below the Commission's "significant modification" marker of roughly a third of original training compute; that is TNW's reading, not a court holding.

NIST remains the US voluntary map. The NIST AI Risk Management Framework was released January 26, 2023, with a generative-AI profile on July 26, 2024. According to NIST, the AI RMF 1.0 is being revised as part of the White House AI Action Plan after that January 26, 2023 release. Map Tenet like any other model that touches client files: what data went in, who can see the output, and who signs the advice.

What this changes in a small legal workflow

Most small firms will not replace Clio this year. They will keep a practice system and bolt a model onto intake, review, and first-draft.

A two-truck shop already photographs a work order and wants a clean job ticket. A solo lawyer already photographs a lease and wants a matter. Clio's Legal Trends hub says firms in its research are meeting the 50%+ of clients who now turn to AI first, growing revenue 4x faster than headcount, and reducing mental strain by up to 25% with technology. The 2026 Legal Trends Report for Solo and Small Law Firms is built on a survey of over 1,700 respondents and warns that if you still bill by the hour, AI time savings can shrink revenue.

The practical stack for a ten-person firm is not "buy Harvey." Keep the system of record, stop pasting privileged text into a consumer tab, and put the model behind a review queue. A shop already routing engagement letters and intake PDFs through US Tech Automations can treat a later legal model as a swap on the extraction step, not as a new matter database.

Transactional shops should watch structured extraction more than the chatbot. Harvey's preview talks about Review Table over up to 10,000 documents with citation quality as the scored object. That is the same job as a closing binder or a vendor-contract dump. Keep the table. Swap the cell model when a specialised checkpoint actually ships.

Family and litigation shops should watch Memory and Spaces. Harvey II says agents inherit the matter, the ethical wall, and the lawyer's style, and that Memory is never used to train models.

If you are still choosing a practice system, start with Clio alternatives for solo lawyers, then the Smokeball and MyCase pieces linked above. Tenet does not score those tools.

Limits you should not round off

Tenet is not in the product yet. Harvey has not published an independent third-party leaderboard for the Tenet checkpoint. Several of the strongest cost claims in the August 20 preview belong to GLM and Qwen post-trains sitting next to Tenet, not to the Kimi K3 checkpoint itself.

The Chinese base is a procurement fact. Open weights help Harvey train. They do not erase client guidelines that bar PRC-origin models, or the Moonshot MaaS licence over $20mn. Harvey's homepage hours-saved figure of 25+ per month is a marketing average for Harvey users, not a Tenet delta.

Signal vs Speculation

Fact, sourced: Harvey announced Tenet as its first in-house legal model on August 18–20, 2026, post-trained on Kimi K3 with Fireworks, and published hold-out lifts on LAB. Harvey's run-rate, valuation talks, customer counts, security posture, and "not live yet" status are all on the record above.

Fact, sourced: Kimi K3 is a 2.8-trillion-parameter open-weight model with a published API price and a MaaS licence threshold that Harvey exceeds. The EU AI Act is in application as of 2 August 2026. NIST's AI RMF remains voluntary.

Our read: if Harvey actually routes a material share of production tokens onto Tenet inside 12 months, small and mid-size firms will not buy Tenet. They will buy the pattern: keep files in one vault, log the model ID on every draft, and let a human sign the letter. Practice-management vendors will offer a "legal-tuned" checkpoint the same way they already offer an OpenAI switch.

Our read: if Tenet stays a research preview through 2027, the story reverts to routing. A small firm should still stop pasting privilege into consumer chat, but it should not wait for a Harvey-class model of its own.

Our read: billing is the SMB tell. If AI cuts draft time and the firm still sells hours, cash falls. Firms that switch those tasks to flat fees will pocket the cost cut even if they never hear the name Tenet. Over 12–36 months, more vertical vendors will hit the same open-weight MaaS licence wall Harvey already clears at 17.5×.

How to act this month

Do not issue an RFP for a foundation model. Pick one document type that already leaves the building: intake forms, executed vendor contracts, or discovery productions.

Put that type on a path you control. A team already routing those files through US Tech Automations workflows will plug a legal-tuned model in as a model swap, not a rebuild, once a vendor actually serves one.

Label the model on the output. Partners cannot supervise a mystery. Revisit pricing on the tasks you are about to speed up.

When you want the agent path rather than another chatbot seat, use the Harvey Tenet workflow map on the US Tech Automations platform to keep extraction, review, and the system of record in one place.

FAQs

What is Harvey Tenet?

Harvey Tenet is Harvey's first in-house legal AI model, a post-trained version of Moonshot's open-weight Kimi K3 built for long-horizon legal tasks. Harvey published the research preview on August 20, 2026.

Is Harvey Tenet live for customers?

No. Business Insider reports Tenet is not live in Harvey yet and that the company will not say when it will be. Treat current Harvey usage figures as the product around Tenet, not as Tenet itself.

Why does a Chinese open-weight base matter to a US solo?

Because the base model is a subprocessor and a licence. Open weights let Harvey train, and they also create a Moonshot agreement above $20mn of MaaS revenue.

Does Tenet train on my client's files?

Harvey's Tenet preview says it did not use customer data in post-training, and Harvey's security page says inputs, outputs, and uploads are not used to train underlying models unless a customer explicitly requests a bespoke model.

What should a 10-person firm change this quarter?

Stop pasting privileged text into consumer chat, put one document type behind a logged extraction-and-review path, and change pricing on any task you are about to make faster. You do not need Harvey's model to do those three things.

Glossary

• Harvey Tenet: Harvey's first in-house legal model, post-trained on Moonshot's Kimi K3.

• Post-training: Extra training on an already-trained base, here with lawyer-written tasks and reinforcement learning.

• Open-weight model: Weights you can download and adapt, still subject to a licence and any MaaS revenue threshold.

• Kimi K3: Moonshot's 2.8-trillion-parameter open-weight model, listed at $3 / $15 per million tokens.

• Legal Agent Benchmark (LAB): Harvey's open-source long-horizon test: 1,200+ tasks, 24 practice areas, 75,000+ rubric criteria.

• All-pass: A LAB task counts only when every rubric criterion passes.

• GPAI: General-purpose AI models under the EU AI Act, in force for those models since August 2025.

Harvey Tenet is a model-ownership play wrapped in a legal product. Small firms will not buy the checkpoint. They can copy the workflow: own the route the file takes, log the model that touched it, and keep a human on the signature line. For the agent-side version of that route, open the agentic workflows board.