Hanno Labs sells bespoke AI engagements to regulated industries and ships its own products, Standd and Hey Lefty. Its open releases include Bosun, a 600M and 4B relevance judge, and three Dinghy Law embedding models. Funding and pricing are undisclosed, so buyers book a call.
Hanno Labs is a US-based private applied AI lab started by Julie Saltman and Stephen Solka. It builds proprietary models and systems for legal, compliance and investment teams, and publishes open models including the Dinghy Law retrieval family, which the lab reports at #1 on MTEB(Law, v1).
Founded: 2025 · HQ: United States · Team: 1-10 · CEO: Julie Saltman
About Hanno Labs
Hanno Labs is an applied AI lab founded in 2025 by Julie Saltman and Stephen Solka, based in the United States. Its homepage describes the company as building proprietary models and systems for highly regulated industries, where legal, compliance, strategy and investment teams need AI that tracks a changing world and still leaves the final call to a person. The lab is private and has not disclosed funding.
Julie Saltman, CEO and co-founder, spent 11 years at the US Department of Justice in regulatory law, served as Assistant General Counsel at the US Copyright Office, and teaches as an adjunct professor at Georgetown Law. Stephen Solka, CTO and co-founder, has 15 or more years building AI and machine learning products at fintech startups and was VP Engineering at AngelList. Research advisors are former Department of Defense scientists who built models to predict adversarial technological surprise for the US Navy.
The commercial side is a services-plus-products model. Hanno Labs says it works problem-backward with clients: define the workflow, build the models and evaluations, then deploy into the client's environment. It lists four platform areas on its site: continuous intelligence, risk monitoring and prediction, institutional knowledge, and decision systems. Its own products are Standd for AI-assisted diligence across complex document sets, Hey Lefty for continuous research and horizon scanning, and Hanno:Biopharma for competitive and regulatory intelligence in biopharma. Pricing is not published; the site routes buyers to a booked call.
The open-source side is where the lab publishes verifiable work. In July 2026 it released the Dinghy Law family of legal retrieval embedding models. By the lab's own leaderboard-faithful reproduction, dinghy-law-8b ranks first on MTEB(Law, v1) by Borda, with dinghy-law-4b third and dinghy-law-0.6b tenth; the lab notes the official board may lag while submissions are reviewed. The 8B model starts from NVIDIA's Nemotron-3-Embed-8B and ships under the OpenMDW 1.1 licence, while the 4B and 0.6B models use Apache-2.0. The weights are hosted on Hugging Face. Other releases include Bosun, a small relevance judge for curating agent memory graphs, released in June 2026 with the WarrantBench benchmark, and langset, a model that embeds text into a geometry you define in a sentence. On 22 September 2026 it published DecisionBench, an open benchmark for models that answer typed questions with probabilities; the first release covers 43 tasks across 28 domains and includes results from Jev, the decision model from TypeSafe, alongside open decision models and general-purpose LLMs. Readers weighing other options in that niche can start with the guide to Jev alternatives.
Where it is not the right fit: Hanno Labs sells bespoke engagements, so a team that wants a self-serve legal assistant it can buy by credit card is better served by a packaged product such as Harvey or Legora. The open models are also narrow tools, not general assistants; teams comparing them with frontier lab offerings from Anthropic, Google or OpenAI should treat them as retrieval and judging components inside a larger system. To compare legal and compliance AI options for a specific job, Smart Match shortlists from the full directory.
Mission
Apply frontier AI to the work that still requires expert judgment
Products
- Dinghy Law: https://huggingface.co/Hanno-Labs/dinghy-law-8b-v1
- Bosun: https://huggingface.co/Hanno-Labs/bosun-4b
- Standd: https://standd.io
- Hey Lefty: https://heylefty.com
- Hanno:Biopharma: https://hannolabs.ai/biopharma
- langset: https://github.com/Hanno-Labs/langset
Hanno Labs Models on Hokai
Links
Frequently Asked Questions
What does Hanno Labs do?
Hanno Labs is an applied AI lab that builds proprietary models and systems for highly regulated industries. Its site names four focus areas: ongoing monitoring, risk prediction, internal knowledge systems and decision support. It sells to legal, compliance, strategy and investment teams, and also runs its own products, Standd, Hey Lefty and Hanno:Biopharma.
Who founded Hanno Labs and who runs it?
Julie Saltman, CEO and co-founder, is a former Justice Department regulatory lawyer who also teaches at Georgetown Law. Technology is led by Stephen Solka, who held the VP Engineering role at AngelList earlier. Research advisors are ex-Defense Department scientists.
What open-source models has Hanno Labs released?
The lab has released the Dinghy Law legal retrieval embedding models in 0.6B, 4B and 8B sizes, the Bosun-XS and Bosun-4B relevance judges with the WarrantBench benchmark, the langset embedding model, and the DecisionBench benchmark. The 8B Dinghy model uses the OpenMDW 1.1 licence, and the 4B and 0.6B models use Apache-2.0, per the lab.
How good is dinghy-law-8b on legal retrieval?
In the lab's own leaderboard-faithful reproduction, dinghy-law-8b scores 72.58 Mean(Task) and tops the MTEB(Law, v1) board on Borda points, 21 points ahead of the next model. Its 4B and 0.6B siblings rank third and tenth. The lab says the official MTEB board may lag while submissions are reviewed, so treat the ranking as vendor-reported.
How much does Hanno Labs cost?
Hanno Labs does not publish prices. Its site sends buyers to a booked call to define a project, and the open models on Hugging Face are free to download under their stated licences. Check each model card for licence terms before commercial use, because the 8B Dinghy model is not Apache-2.0.