Glossary
Open-Weight Model — Definition for Lawyers
An open-weight model publishes its trained weights for anyone to run. Open weights vs open source, DeepSeek, Mistral and Muse Glimmer licences, self-hosting and government bans.
On this page
An open-weight model is an AI model whose trained parameters, the weights, are published for anyone to download and run on their own hardware, under a licence such as MIT or Apache 2.0. It is not the same as open source: the training data and the training code stay private. For law firms the attraction is control over where client data goes; the price is a measurable quality gap.
Open weights, open source and closed: three different things
A model is a very large file of numbers plus the code that runs them. Publish the numbers and you have an open-weight model. Publish the numbers, the code and the data the model was trained on, under a licence that lets others reproduce and modify all of it, and you have something close to open source in the software sense. Keep the numbers on your own servers and sell access through an API, and the model is closed. The general-purpose AI model entry explains the EU AI Act's separate vocabulary for the same objects.
| Open-weight | Open source (full) | Closed | |
|---|---|---|---|
| Weights downloadable | Yes | Yes | No |
| Training data published | Usually not | Yes | No |
| Runs on your own hardware | Yes | Yes | No |
| Licence | MIT, Apache 2.0, or a custom vendor licence with restrictions | OSI-style, covering data and code | Terms of service |
| Examples, September 2026 | DeepSeek V4 (MIT), Mistral Large 3 (Apache 2.0), Meta Muse Glimmer (Apache 2.0), Nvidia Nemotron 3 | Rare at frontier scale | Claude Fable 5.1 and Opus 5, GPT-6 Astra, Gemini 3.1 Pro, Meta Muse Spark |
The licence line deserves reading rather than assuming. DeepSeek V4-Pro, a 1.6-trillion-parameter model released on 13 August 2026, ships under the MIT licence, about as permissive as licences get. Mistral Large 3, a 675-billion-parameter mixture-of-experts model from December 2025, uses Apache 2.0. Meta's Muse Glimmer, 30 billion parameters and small enough to run on a 24GB GPU, shipped under Apache 2.0 on 10 August 2026, while the weights for the far stronger Muse Spark were promised "soon" on 2 September 2026 and, as of this page, have not appeared. Harvey describes Tenet, its post-trained legal model in research preview since 20 August 2026, as open-weight, but the terms on which a non-customer could obtain it were not published in the sources for this page. "Open-weight" on a press release is a claim to verify like any other.
Why a law firm would want one
Three reasons, in descending order of how often they apply.
Confidentiality that does not depend on a contract. A model running on your own server has no vendor retention period, no abuse-monitoring log, no subprocessor and nothing for a third party to subpoena from the provider. That is stronger than a zero data retention clause, and it is why Latham & Watkins bought Nvidia H200 servers housed where only its staff can go and is fine-tuning Nemotron 3 models on them, according to reports of 10 and 11 September 2026. The CCBE's Technical Guide on AI tools and models, adopted 27 March 2026, walks through deployment models from on-premise to SaaS and treats open-weight models as a legitimate option for exactly this reason.
Continuity. On 12 June 2026 the US government restricted access to Claude Fable 5 for foreign nationals, and Anthropic suspended every user's access worldwide for 18 days because it "had no reliable way to verify nationality in real-time". A model on your own hardware does not get switched off by someone else's regulator in the middle of a matter.
Sovereignty. Mistral's pitch to European firms is that an EU-domiciled provider matters more than a data-residency toggle, because a toggle changes where data sits, not which government can compel its production. Self-hosting takes that argument to its conclusion.
The costs: quality, bans and staff
The quality gap is the number to have in front of you before any of the above persuades you. On Vals' Legal Research Bench, DeepSeek V4-Pro scored 40.87%, and on Harvey's strict Legal Agent Benchmark 7.50%; the top of those leaderboards on 15 September 2026 was Meta's Muse Spark 1.3 Max, whose weights remain unpublished, with Claude Opus 5 and Fable 5.1 behind it. In HAQQ's June 2026 test, Mistral Large produced citation errors on 64% of questions. A firm choosing an open-weight model for legal analysis is choosing control over capability, and should say so in its risk register.
Origin follows the weights. The FY2026 NDAA orders DeepSeek removed from US Department of Defense and intelligence systems, and Australia banned it on government systems in February 2025. A government or defence client's guidelines will apply whatever server room the model runs in. The staff cost is real too: Latham has about 100 technology people on AI out of 900. At the other end of the scale, Muse Glimmer runs on a single 24GB GPU, and the open-weight versus closed comparison sets out what sits between a data centre and a desktop card.
Five questions before self-hosting
- Which matters genuinely cannot go to a frontier model on enterprise terms with a signed data-processing agreement? If the honest answer is "a handful a year", price the project against that handful.
- Which licence, read in full, and does it permit commercial use and modification?
- Who patches, monitors and evaluates the model, and what happens when they leave?
- What is the accuracy on your own test set of lawyer-marked questions, against Opus 5 or GPT-6 Astra on the same set?
- Does any client's policy ban the model's country of origin?
The build-versus-buy guide turns those into a decision tree by firm size. The AI Lab for Lawyers course stays with browser-based tools throughout, and its final module on risks and regulation is where the self-hosting question gets its hearing.
Frequently asked questions
What is the difference between an open-weight model and an open-source model?
An open-weight model publishes its trained parameters so anyone can download and run it, usually under MIT or Apache 2.0, but keeps the training data and often the training code private. A fully open-source model would publish all three under a licence that allows reproduction and modification. At frontier scale almost nothing is fully open source; DeepSeek V4, Mistral Large 3 and Meta's Muse Glimmer are open-weight. A closed model such as Claude Opus 5 or GPT-6 Astra is reachable only through its vendor's service.
Are open-weight AI models good enough for legal work?
Measurably behind the frontier as of September 2026. DeepSeek V4-Pro scored 40.87% on Vals' Legal Research Bench and 7.50% on Harvey's strict Legal Agent Benchmark, while the top of both leaderboards on 15 September 2026 was Meta's Muse Spark 1.3 Max, whose weights are unpublished, followed by Claude Opus 5 and Fable 5.1. Mistral Large produced citation errors on 64% of questions in HAQQ's June 2026 test. A firm that self-hosts is choosing control over capability, which is a legitimate choice for a narrow class of matters.
Why would a law firm self-host an open-weight model?
For matters that cannot leave the building. A model on the firm's own hardware has no vendor retention period, no abuse-monitoring log, no subprocessor and nothing a third party can subpoena from the provider. Latham & Watkins bought Nvidia H200 servers held where only its staff have access and fine-tunes Nemotron 3 models on them, per reports of 10 and 11 September 2026, while still using Harvey, Legora and frontier tools for everything else. The CCBE's March 2026 Technical Guide treats on-premise open-weight deployment as a recognised option.
Put this into practice
AI Lab for Lawyers: a live, hands-on course for practising lawyers
4 live sessions (8 hours), recordings included, taught by Dr. Niklas Schmidt of Wolf Theiss. Rated 4.7/5 from 381 ratings; alumni from Norton Rose Fulbright, Paul Hastings, Pillsbury and Taylor Wessing.
Related reading
Comparisons
Open-Weight vs Closed Models for Law Firms: Control, Cost and Confidentiality
Open-weight vs closed AI models for law firms: what Latham's own GPUs buy, Harvey Tenet, DeepSeek V4, Mistral, the legal benchmark gap, when self-hosting pays.
10 min read
Guides
Build or Buy Legal AI? Latham's GPUs, Harvey's Seats and the Firms in Between
Build vs buy legal AI in 2026: rent Harvey or Legora, assemble on Claude or Astra with MCP, or build like Latham on own GPUs. Costs, control, decision tree.
11 min read
Glossary
Fine-Tuning — Definition for Lawyers
Fine-tuning retrains an AI model's weights on your data. How it differs from RAG and prompting, why Latham and Harvey do it, why Spellbook refuses, and the client-data problem.
5 min read
Glossary
Zero Data Retention (ZDR) — Definition for Lawyers
Zero data retention means the AI vendor stores nothing after replying. Which plans offer ZDR, the Covered Models exception, and why it matters for privilege and GDPR.
5 min read
Glossary
General-Purpose AI Model (GPAI) — Definition for Lawyers
A general-purpose AI model under the EU AI Act is a model with significant generality, such as GPT-6 or Claude Opus 5. Definition, tiers, signatories, enforcement.
4 min read
Regulation & Ethics
AI Vendor Due Diligence for Law Firms: The Questions to Ask Before You Sign
AI vendor due diligence for law firms starts with the subprocessor chain, then training and retention terms, ZDR eligibility, data residency, audit logs and exit rights.
11 min read