Harvey Tenet on Kimi K3: What the Legal Post-Train Really Shows

Harvey Tenet is a legal-domain post-train of Kimi K3, not a new Kimi model. Harvey, the legal AI company, took Moonshot’s open-weight K3 base and post-trained it with Fireworks using reinforcement learning on long-horizon legal agent tasks. The result nearly doubled K3’s all-pass rate on Harvey’s held-out Legal Agent Benchmark (10.8% to 19.7%) at roughly the same cost per task. Tenet is a research preview, and you cannot download it or call it through an API.

This article covers what is confirmed as of September 25, 2026, five weeks after Harvey’s August 20 announcement: the benchmarks in absolute terms, the three cost claims, why Harvey chose an open-weight base, data residency, and lessons for anyone fine-tuning K3. Where we couldn’t verify something, we say so.

What Harvey actually announced

The primary sources are Harvey’s own post, Harvey Tenet Research Preview (August 20, by Calvin Qi, Vasudha Rengarajan, Julio Pereyra, Niko Grupen and Gabe Pereyra), and Fireworks’ companion write-up, Post-training Kimi K3 with Harvey for long-horizon legal work (August 26). Harvey frames the work around two research goals: “building frontier legal intelligence using open-weight models” and “creating systems to allow law firms to build their own specialized models and own their intelligence.”

Item	What the sources confirm
Announced	August 20, 2026 (Harvey blog + X post)
Status	Research preview. No weights, model card, API or release date
Base model	Kimi K3 (2.8T-parameter MoE, open weights, July 2026)
Training partner	Fireworks, via asynchronous RL on the Fireworks Training API
Algorithm	Group-Sequence Policy Optimization (GSPO)
Adapter	Rank-64 LoRA over the full K3 network
Compute	~150 NVIDIA B300 GPUs for about 2 months
Environments	~1,750 agentic legal task environments
Rollouts	8 task groups × 8 rollouts per optimizer step; 10,000+ rollouts per epoch
Rollout length	Harvey: some exceed 1,000 turns and hundreds of thousands of tokens. Fireworks: each runs past 50 turns and 100,000 generated tokens
Grader	”Kimi 2.6”, in Harvey’s wording, used as the judge
Training data	Synthetic, publicly available legal data, human expert data. “We did not use any customer data”
Data partners	Mercor and others built the human expert datasets

Two details are easy to miss. First, this is a LoRA post-train, not a full-weight retrain. A rank-64 adapter sits on top of the unchanged K3 base, so serving Tenet still means serving all of K3. Second, Harvey’s post covers more than Tenet. It also describes specialist work on other open models: a GLM-5.2 orchestrator for M&A diligence, a post-trained model for high-volume “Review Table” extraction, and a Qwen-based model for firm-knowledge search. Some summaries mix those numbers into Tenet’s. They are separate results.

The benchmark numbers, in absolute terms

The headline “almost twice as many tasks” is accurate. It also starts from a low base, so the absolute gains are worth reading alongside it. The table uses the figures Fireworks published. The relative column is our arithmetic.

Benchmark	Base Kimi K3	Harvey Tenet	Absolute change	Relative change
LAB all-pass (held-out tasks)	10.8%	19.7%	+8.9 pts	+82%
LAB Contracts all-pass	9.3%	11.3%	+2.0 pts	+22%
Mercor Apex Agents (Corporate Law)	58.8%	74.0%	+15.2 pts	+26%
Crosby Redline Bench	49.3%	55.5%	+6.2 pts	+13%
LegalBench, CUAD, MAUD, Apex-v1	baseline	”holds its performance”	not published	not published

Source: Fireworks and Harvey. All results are vendor-run.

How to read these:

• “All-pass” is a strict metric. A task counts only if every rubric criterion is met. So Tenet fully completed about one in five held-out tasks, and four in five still missed at least one criterion.
• The transfer results matter most. Apex Agents and Redline Bench were not in Tenet’s training environments. Gains there (+15.2 and +6.2 points) suggest the RL taught general long-horizon legal agent behavior and did not just overfit to Harvey’s own tasks.
• Knowledge benchmarks stayed flat, as intended. Harvey says Tenet “maintains strong performance” on LegalBench, CUAD, MAUD, Mercor’s Apex-v1 and Scale’s Professional Reasoning Benchmark. No before/after numbers were published.
• LAB Contracts moved least. Harvey calls Tenet “state-of-the-art” there, but the gain was 2 points.

Three leaderboards, three different K3 scores

This part confuses a lot of coverage. At least three versions of “Harvey’s Legal Agent Benchmark” are in circulation, and K3’s score depends on which one you read.

Version	Who runs it	Kimi K3 score	Context
Harvey’s held-out LAB (Tenet paper)	Harvey	10.8% all-pass	Tenet: 19.7%
Vals AI “Harvey’s Legal Agent Benchmark”	Vals AI	10.83% (rank 12)	Top: Muse Spark 1.2 at 25.42% (BenchLM mirror, Sep 22 snapshot)
Harvey LAB-AA	Artificial Analysis	26.7% all-pass	Muse Spark 1.3 30.8%, Claude Fable 5.1 16.7%, Grok 4.6 15.8% (BenchmarkList, Sep 4)

Base K3’s 10.8% in the Tenet write-up almost exactly matches its 10.83% on the Vals leaderboard. That suggests a similar harness, but Harvey doesn’t say so, and “held-out tasks” could be a different subset. Don’t compare Tenet’s 19.7% with LAB-AA scores. LAB-AA uses a different implementation, where base K3 already scores 26.7%.

One more conflict: Kimi’s own legal-work resource page (updated September 16) calls K3’s 26.7% “the highest result in that evaluation.” The Artificial Analysis snapshot we found lists Muse Spark 1.3 (xhigh) ahead at 30.8%. Leaderboards change often, so the claim may have been true when it was written. Check the live board before you repeat it. Together AI separately claimed K3 scored “roughly 60% higher” than Claude Fable 5.1 on LAB-AA’s hard subset, but it published no absolute hard-split scores or task counts.

For a broader comparison with Anthropic’s model, see our Kimi K3 vs Claude Fable 5.1 breakdown.

The cost story: three claims measuring different things

Claim	Source	What it compares	How solid
$5.92 per LAB task vs $5.62	Fireworks blog	Tenet vs base K3, per attempted task	Specific, vendor-measured
”Less than a fourth the cost of leading foundation models”	Harvey’s X announcement	Tenet vs unnamed closed models	Comparators and method not disclosed
Gross margin from ~50% to -50%, then positive after August	Bloomberg, Sep 21	Harvey’s whole business	Reported, not itemized

Per attempted task, Tenet is about 5% more expensive than base K3 ($5.92 vs $5.62). It completes far more tasks, though, so the cost per fully completed task falls sharply. Our rough arithmetic, assuming both figures cover the same held-out set: $5.62 ÷ 10.8% ≈ $52 per all-pass task for base K3, versus $5.92 ÷ 19.7% ≈ $30 for Tenet. That is about 42% cheaper per completed task. Harvey says it got there with reward shaping that favored trajectories using fewer tokens “given equivalent performance.” Its post puts it this way: “cost is a function of both token prices and tokens used.”

The “less than a fourth” figure compares Tenet with unnamed closed models and mixes K3’s lower token prices (Moonshot lists $3 input / $15 output per million, $0.30 cached) with Tenet’s token efficiency. Until Harvey names the comparators, treat it as marketing.

The Bloomberg figures are the most important for the business side. According to Bloomberg’s September 21 report (syndicated on Yahoo Finance), token usage rose twentyfold after a March agent update, and gross margins went from about 50% at the start of 2026 to about -50% by June. Margins turned positive again after the August model launch “and other changes” to how Harvey uses AI. The same report says Harvey still uses Anthropic’s Claude Opus for the hardest tasks. Tenet did not replace frontier models across the board, and nobody has published how Harvey’s workload splits between them.

Why an open-weight base, and why K3

Harvey’s posts give no model-by-model reason for picking K3 over other open bases. What the sources support:

1. Ownership of the model. Harvey’s stated goal is to let law firms “own their intelligence.” In a Sequoia Capital podcast from August (transcribed by BigGo Finance), co-founder Gabe Pereyra said: “The end game here is not for us to build the best legal model. It’s for us to help every law firm or enterprise customer customize it to the type of work they’re doing.” You can’t give a customer a model you only rent through an API.
2. Open models are competitive on narrow tasks. From the same interview: “Maybe not general frontier intelligence, but if you have a specific task like us, they are competitive.” The LAB-AA leaderboard supports this for K3 in particular, with base K3 well ahead of Claude Fable 5.1 there.
3. Unit economics. Bloomberg’s margin figures explain the timing. A company paying usage-based prices on twentyfold token growth has a strong reason to own the model.
4. Scale and context. K3 is the largest open-weight model available (2.8T total parameters, 104B active, 1,048,576-token context; see our specifications page). Long-horizon legal tasks that run for hundreds of turns benefit from a 1M window and strong agent behavior.

SCMP called the move “a significant departure” for a company backed by the OpenAI Startup Fund, Sequoia and Andreessen Horowitz that had built on closed models from Anthropic, OpenAI and Google. Harvey has kept growing since: Tech Times reported a $550 million round at a $15.6 billion valuation on September 9, with about $400 million in ARR.

Deployment, data residency and the “Chinese base” question

What open weights make possible. K3’s weights are downloadable (~1.56TB, see our open weights guide). A post-trained K3 can run on any infrastructure its owner controls, and no requests go to Moonshot. Several outlets describe this as the core privilege argument: client documents stay with the firm’s chosen provider instead of a model lab. The Next Web put it this way: “Harvey has quietly turned it into a question about who your counterparty is when the work is covered by privilege.”

What has not been disclosed. Harvey has not said where Tenet will run in production, which provider will serve it, or which regions. The Tech Times headline about keeping “client documents in firm control” describes the architecture open weights allow. It is not a confirmed deployment spec for Tenet. Until Harvey publishes serving terms, data residency for Tenet is unknown.

The regulatory question. The Next Web points out that under the EU AI Act, a company that modifies a general-purpose model becomes its “provider” only if the changes are substantial. So “most of the obligations stay upstream,” and a European firm using Tenet would rely on “a documentation chain that begins in Beijing.” Bloomberg also reports that Rogo, another startup exploring its own model training, has had to talk with customers who are hesitant about Chinese models. For procurement teams, the base model’s origin is a governance question even if no data ever reaches China.

For teams using K3 directly today, the residency options are the same as for any open-weight deployment: US-hosted providers such as Fireworks (which offers a zero-data-retention endpoint), EU hosting through Nebius, or self-hosting. Our API providers guide covers the options.

What Tenet signals for domain post-training on K3

Tenet is the first public, benchmarked example of a company post-training K3 for a vertical. Four points are useful beyond the legal field.

1. The tooling is available to anyone. Fireworks announced K3 fine-tuning (supervised fine-tuning, preference tuning and RL) through its Training API. Its K3 LoRA post (July 26) describes serverless LoRA training billed per token, dedicated capacity for full-parameter runs, and multi-LoRA serving. It gives one example: a small RL run of about 20 steps and 860K training tokens cost “about $65 all in.” The Training API became generally available on August 31 and cites Tenet as a case study.

2. Environments and data are the real cost. Harvey’s compute (about 150 B300s for two months) is large but rentable. The ~1,750 graded environments, expert rubrics and hired-lawyer datasets are much harder to copy. If you don’t have realistic, gradable tasks in your domain, RL on K3 has little to learn from.

3. A LoRA post-train still needs the full base model at inference. Rank-64 adapters are small, but they only work on top of all of K3. Self-hosting means ~1.7–1.84TB of GPU memory (8×B300 or 8×MI355X-class at minimum; 8×H100 cannot fit it). Details are in our VRAM requirements guide. Most teams will serve adapters on a provider that supports multi-LoRA instead.

4. Read the license before you build a business on it. The Kimi K3 License has two clauses that matter for a Harvey-style company:

Clause	Trigger	Requirement
Model-as-a-Service	MaaS business with licensee + affiliates’ revenue above $20M over any 12 consecutive months	Separate agreement with Moonshot before commercial use of the software or derivatives
Attribution	Commercial product above 100M MAU or $20M monthly revenue	”Kimi K3” prominently displayed in the UI
Exemptions	Internal use; access via Moonshot’s official products or certified inference partners	Above requirements don’t apply

The license’s MaaS definition covers API access where users control inputs and parameters, and excludes embedded features. A legal assistant that embeds Tenet in its product is not the same as reselling raw model access. Whether Harvey’s arrangement falls under either clause has not been published. We cover the license in detail in our Kimi K3 license explainer.

What it means for you

You are…	What Tenet changes	What to do now
A lawyer or firm using Harvey	Possibly better long-horizon agent work later; Tenet is not in the product yet	Ask Harvey which model runs which workflow, where it’s hosted, and its retention terms
A firm evaluating AI vendors	”Built on a Chinese open-weight base” is now a real procurement question	Ask about serving location, provider chain, and EU AI Act documentation
A legal-tech builder	Evidence that RL on K3 improves agentic legal work	Build gradable task environments first; try base K3 with a strong harness before paying for RL
An enterprise considering K3 fine-tuning	A working template: LoRA + GSPO + expert-graded environments	Start with a small serverless LoRA run; check the $20M MaaS clause if you’ll resell access
A casual Kimi user	Nothing. Tenet isn’t on kimi.com or the Kimi app	Use base K3 as usual (our complete guide covers the options)

Builders should test base K3 on their own tasks first; on LAB-AA it already beats several closed models. Harvey’s biggest gains came on long, multi-step agent tasks, and the smallest on LAB Contracts. For short Q&A or clause extraction, a good harness and retrieval may be enough.

What we could not verify

• Tenet’s production rollout. Bloomberg ties Harvey’s margin recovery to the August launch “and other changes”, but no source says how much traffic Tenet (or any K3-based model) serves.
• The “less than a fourth” cost claim. It appears in Harvey’s X announcement with no named comparators.
• Training cost in dollars. Only GPU count and duration are public.
• Moonshot’s involvement. We found no Moonshot statement about Tenet beyond Kimi’s resource page mentioning that Tenet “outperformed today’s leading models across several legal benchmarks.”
• A widely quoted Pereyra line about model labs being a “future competitor”, and a figure of 13 trillion tokens per month. Both appear in secondary coverage without a primary source we could check, so we left them out.

FAQ

What is Harvey Tenet? Harvey Tenet is Harvey’s first post-trained model: Kimi K3 as the base, post-trained with Fireworks using asynchronous reinforcement learning on long-horizon legal agent tasks. Harvey announced it on August 20, 2026 as a research preview. It is not a Moonshot product and not a new Kimi release.

Can I download or use Harvey Tenet? No. As of September 25, 2026, Harvey has not published weights, a model card, an API endpoint or a release date. Harvey says it is working to bring its research into production inside the Harvey product, but Tenet itself is not available to download, license or call directly.

How much better is Tenet than base Kimi K3? On Harvey’s held-out Legal Agent Benchmark (LAB) tasks, the all-pass rate rose from 10.8% to 19.7% (+8.9 points, about 82% relative). On LAB Contracts it rose from 9.3% to 11.3%. On two outside benchmarks it was not trained on, Mercor’s Apex Agents (Corporate Law) went from 58.8% to 74.0% and Crosby’s Redline Bench from 49.3% to 55.5%. All of these numbers are Harvey- and Fireworks-reported.

Was Tenet trained on client data? Harvey says no. Its post states that no customer data was used in any of its post-training work. The training mix was synthetic data, publicly available legal data and human expert data, with Mercor among the partners that helped build expert datasets. That is Harvey’s own statement, not an independent audit.

Does using a Kimi K3 base mean data goes to China? Not by itself. K3 is open-weight, so a post-trained version can be served on any infrastructure its owner chooses, with no calls to Moonshot’s API. Harvey has not published where Tenet will run in production, so treat data residency as an open question until it does. The weights’ origin can still matter for procurement and regulatory reviews.

How much did it cost to train Tenet? Harvey has not published a dollar figure. It says the run used about 150 NVIDIA B300 GPUs over two months, roughly 1,750 agentic legal environments and more than 10,000 rollouts per epoch, with a rank-64 LoRA over the full K3 network. On inference, Fireworks reports $5.92 per LAB task for Tenet versus $5.62 for base K3.

Can my company fine-tune Kimi K3 the way Harvey did? The tooling is public. Fireworks offers K3 fine-tuning through its Training API, including LoRA training and reinforcement learning on serverless or dedicated capacity. The hard part is the 1,750 graded task environments and expert data, not the GPUs. Also check the Kimi K3 License: model-as-a-service businesses above $20M in revenue over 12 months need a separate agreement with Moonshot.

Is Kimi K3 good at legal work without post-training? It is competitive. On Artificial Analysis’s version of Harvey’s benchmark (LAB-AA), K3 scored a 26.7% all-pass rate as of early September 2026, second only to Muse Spark 1.3 (30.8%) and ahead of Claude Fable 5.1 (16.7%). On the Vals-hosted LAB leaderboard it sits at 10.83%. The two versions use different setups, so compare within one leaderboard only.

The bottom line

Tenet shows that K3 can be post-trained into a stronger domain agent at a reasonable cost. Harvey nearly doubled fully completed held-out legal tasks (10.8% to 19.7%), improved on two outside benchmarks it didn’t train on, and kept per-task cost roughly flat. The limits matter as much as the results. Tenet is a research preview, the numbers are vendor-run, the absolute pass rates are still low, and Harvey reportedly still uses Claude Opus for its hardest work. The broader takeaway is for everyone else: the tools to do this on K3 are public, and the real advantage is in the task environments and expert data.

Keep reading: Kimi K3 open weights, Kimi K3 license explained, VRAM and hardware requirements, every K3 API provider, and Kimi K3 vs Claude Fable 5.1.