What Post-Training an Open-Weight Model Actually Buys a Company
- David Borish
- 1 day ago
- 5 min read
Harvey spent two months and roughly 150 Nvidia B300 GPUs turning a Chinese open-weight model into what it calls the best available system for contract review. The legal AI company introduced Tenet on August 20, a model built on Moonshot AI's Kimi K3, a 2.8-trillion-parameter base released with downloadable weights. Harvey did not build a new foundation model. It took someone else's and taught it to practice law.
That distinction, foundation model versus post-trained model, is becoming the central decision enterprises face in 2026. Harvey's own numbers show why the choice increasingly favors post-training.
What Tenet Actually Learned
Harvey trained Tenet using asynchronous reinforcement learning on a corpus that combined publicly available legal data, synthetic tasks, and work reviewed by practicing attorneys. According to Harvey's research preview, the model completes almost twice as many held-out tasks on the company's Legal Agent Benchmark as the untrained Kimi K3 base, and 20 percent more on the contracts-focused version of that benchmark. Harvey frames those as an 82 percent and 22 percent relative improvement in all-pass rate. In absolute terms, that's a gain of roughly nine percentage points on the main benchmark and two points on contracts, which matters for reading the claim correctly: Tenet's edge is real, but the underlying pass rates it's moving from and to are still modest in absolute terms.
Harvey reports state-of-the-art results on the contracts benchmark and second place overall, and says the gains transferred to benchmarks it never trained on, including Mercor's Apex Agents corporate law track, Crosby's Redline Bench, and Scale AI's Professional Reasoning Bench. The company also says Tenet holds its footing on benchmarks that test static legal knowledge rather than agentic task completion, meaning the specialization didn't come at the expense of general competence.
The mechanics behind that gain are the real story for other companies weighing the same path. Harvey's team, working with Fireworks AI, optimized the model using group-sequence policy optimization, adjusting a rank-64 LoRA layer across the full Kimi K3 network rather than retraining the whole thing from scratch. Each training task dropped the model into a simulated legal matter, complete with documents, an instruction written the way a partner would actually phrase a request, and a rubric of dozens of specific criteria a finished work product had to satisfy. A separate model graded each attempt. Harvey ran more than 10,000 rollouts per training epoch this way over roughly 1,750 task environments.
Fireworks solved a harder problem underneath that: keeping a massive mixture-of-experts model numerically consistent between the version generating training rollouts and the version being graded, a mismatch that can quietly wreck reinforcement learning at this scale. Harvey and Fireworks addressed it with matched kernels and a token-in-token-out replay system, a detail that explains why this kind of post-training has been out of reach for most companies until recently. It requires infrastructure most legal, healthcare, or financial services firms don't have in-house.
The Cost Case Companies Actually Care About
Performance is only half of why enterprises are choosing this route. Harvey says the second finding from its research is cost. Open-weight models already carry cheaper per-token prices than closed frontier systems. Harvey layered a second cost lever on top by rewarding the model, during training, for reaching the same quality with fewer tokens. The company reports Tenet now runs at under a fourth the cost of leading foundation models, a figure Harvey has not published independently, so it should be read as a vendor claim rather than an audited benchmark.
That combination, better task completion plus lower and more predictable cost, is what has made post-training attractive beyond legal work. Harvey didn't stop at one general model. It also post-trained three narrower specialist models that Tenet can call as subagents: one for M&A due diligence built with Baseten using a harness that lets a root agent delegate document review across tasks spanning up to 80 million tokens, one for high-volume contract review tables built with Applied Compute that Harvey says improved citation accuracy by 12.1 points at roughly a tenth of the per-cell cost, and one for firm-specific knowledge built with Engram that cut total inference tokens by 58 percent by teaching the model to internalize a firm's own precedent rather than re-searching it on every query. Each of those is a separate, narrower post-training run aimed at a specific workflow rather than a single model trying to do everything.
Harvey Isn't the Only One Doing This
The same pattern is showing up across other verticals. Intercom's customer service model, Fin Apex 1.0, is built on an undisclosed open-weight foundation and, according to the company, resolves customer issues at a 73.1 percent rate against 71.1 percent for GPT-5.4 and 69.6 percent for Claude Sonnet 4.6, while running at roughly a fifth of the cost of deploying those frontier models directly. Fastino Labs took a similar approach in the other direction, publicly releasing specialized finance and healthcare models post-trained on Nvidia's Nemotron 3.5 Lightning base and putting the fine-tuning process itself into an agent other companies can run. JetBrains has done the same for code completion with its Mellum models, publishing intermediate training checkpoints so other teams can see exactly where the specialization happened.
The throughline across all of these is that the differentiation is shifting away from which frontier lab a company rents from and toward what a company trains on top of an open base. As more vertical players get good at post-training, the willingness to pay a premium for the single best frontier model narrows, because a smaller open model tuned on the right data can match or beat it on the tasks that actually matter to that business.
Where This Fits the Open-Prem Curve
This is precisely the shift I mapped in the April 2026 edition of the Open-Prem Inflection Point. That paper documented at least nine distinct open-source model families now operating at or near frontier performance, and found that self-hosted inference runs between five and twenty cents per million tokens against three to fifteen dollars for proprietary cloud APIs, with organizations processing more than two million tokens a day reaching hardware payback in six to twelve months. Harvey's Tenet is what that inflection point looks like inside a single company: an enterprise with enough scale and enough proprietary process data chooses to own the model rather than lease it, because the economics now support that choice and the compliance picture, keeping client data off a third party's servers, is cleaner besides.
There's a second layer to the Tenet story worth naming directly. Kimi K3 is a Chinese-built model, and an American legal AI company backed by OpenAI just put it underneath work covered by attorney-client privilege. I flagged this trajectory back in July 2024, in China's Recent AI Surge Challenges US Dominance: A Wake-Up Call for the West, when Chinese open models climbing international benchmarks was still treated as a curiosity by most of the industry.
What to Watch Next
Harvey says its next steps are expanding LAB to more jurisdictions and practice areas and scaling the compute behind this research into new generalist models for its product. The company has not published Tenet's weights, a standalone license, or an API, so for now this remains a research preview rather than a deployable product outside Harvey's own platform. The open question for every company reading Harvey's research is not whether post-training works. Harvey, Intercom, and Fastino Labs have each shown, in their own words, that it does. The open question is how much proprietary process data and infrastructure a company needs before post-training its own base model beats paying by the token for someone else's.
About the Author
David Borish is an Enterprise AI Strategist and the author of The Tony Hawk Paradox, which argues that capabilities proven in controlled or simulated environments go on to reshape broader systems once the constraints that contained them fall away. He writes on frontier AI, enterprise infrastructure, and open-source deployment economics at davidborish.com.
