Sadly, local models are not quite there yet for legal work

Professor evaluates outputs from a local model, Lexis Protege and Claude grounded with CourtListener

A model that runs on your own laptop never leaks a client's name, never sends an invoice, and never gets discontinued. Which is exactly why I keep testing local models — and exactly why this post is going to disappoint. But you should read it anyway. By the way, it also suggests some issues with Lexis Protege.

Most of us like working on machines we don't own. Someone else buys the chips, patches the server, and swaps in the newest model while you sleep. You just open a browser tab or a lovely desktop app and one of the best systems in the world is waiting. I am not trying to talk anyone out of that. The trouble is the bill that comes with the convenience. For practicing lawyers it arrives in three forms, and law students training for practice should learn to recognize each one. The first is confidentiality: every prompt carries a client's facts to a computer you do not run. The second is money, which is likely to be an ever-growing concern. The third is dependency: the vendor decides what exists. Models are retired on the vendor's schedule, not yours. Claude Fable wondrously arrived, for example, but following pressure from the Trump administration, Anthropic retired it within days of release. You get no vote on whether the tool you built around is still there next week.

Money is the part that was supposed to take care of itself. A token was going to keep getting cheaper until the price stopped mattering. It hasn't at least not for the work lawyers do — legal research means long documents, lots of reasoning tokens and many runs. The better the model, the more each run costs. AI has become a victim of its own success: the demand for its wonders is enormous, and the chips and data centers that feed it cannot be built out fast enough to keep pace. Adam Smith explained the result of this scenario 250 years ago and little has changed since. The direction of travel is an unfortunate one for legal work. The newest models reason in long chains and call tools as they go, spending more tokens per question, not fewer. What has hidden this so far for many users, particularly students, legal academics and small practitioners, is the monthly subscription pricing model — a flat fee that folds in a ration of tokens the vendor has been quietly selling below cost. That arrangement is unlikely to survive rising usage; the tokens bundled into a subscription today are the easy thing to pare back tomorrow, shunted out to the à la carte API calls that bill by the token. A solo watching an expensive model spin the meter through a single coverage question is right to ask whether she's renting something she can never afford to leave running.

Just as I was about the press the "publish" button on this post, the news fortified my concerns that we are moving from subscription models that may subsidize the intensive user to a per token model. Legora, one of the two leading vendors of legal tech to BigLaw, announced that it was migrating to "consumption based pricing" at least as an option. You can read about it here. Many suspect that Legora's move is just the beginning. You can read an interesting commentary by blogger Matt Pollins here. If these changes in pricing models persist, interest in local models that actually work in legal is going to increase.

There are, to be fair, ways to soften these problems short of bringing the model home. A whole market has grown up around the confidentiality worry. Harvey, Legora, Clio, and other platforms are built to wall a firm's data off from outsiders and, just as important, from the wrong people inside the firm; “AI governance” is the current term. It works, more or less, but it is not cheap. The other route is an enterprise contract straight from a vendor, which can buy better protection on paper. Both carry their own version of the dependency problem — you are now locked to a platform or a contract — and neither scales gracefully downward: the price and overhead can make sense for a firm of a certain size, perhaps less for a solo, a public-interest organization, or an individual, who are left roughly where they started.

A model that runs entirely on your own laptop would seem to answer all three concerns at once. The client's facts never leave the machine. So, unless someone subpoenas your logs, the communications are private. No meter runs. And nobody can retire it out from under you. So the only little question left is whether it's any good. The way to find out is not to read benchmarks but to hand it real work and grade what comes back.

I gave it a genuine problem: the duty of a Texas liability insurer to settle, with the wrinkle that makes the doctrine hard — what happens when nobody yet knows whether, at the time a settlement demand is in play, the policy even covers the lawsuit. Then I ran the same question through a frontier cloud model, Claude Opus 4.8, wired as I have recommended, to CourtListener and through Lexis's Protégé. I compared all three. One caveat before the results. The model I tested is about the largest and best that fits on the kind of laptop a practicing lawyer actually owns; I happen to have a fairly high-end Mac, so I'm giving local models their best honest showing, not a crippled one. There's a frontier above this — the 500-gigabyte-plus DeepSeek, Kimi or GLM models that want a $50,000+ refrigerated box in a closet with support staff. Those big models might yet change the verdict for those willing to incur the charges and to trust models developed in a nation that some mistrust as an adversary. Examination of that alternative, however, is for another day.

The Local Model Experiment

Here's the report of my experiment: TL;DR Unless you can build some sort of harness that does a better job than me, don't rely yet on local models for serious legal work.

Using the free LM Studio product, I loaded a good size local model into my Mac laptop: google/gemma-4-26b-a4b-qat. I used that model because it is a "reasoning model" and because it can use "tools" such as an MCP connector to CourtListener. It's about the minimal model I thought might have a decent chance. I'd tried smaller ones earlier and they were completely hopeless. (Some bigger ones from Meta and Mistral were hopeless too). I gave Gemma a question about Texas insurance law, an area with which I am familiar, so that I could augment any AI evaluation of the answer with my own domain knowledge.

Research the duty to settle of liability insurers under Texas law with particular attention to special procedural and doctrinal issues that arise when it is unclear whether the policy covers the event giving rise to the lawsuit.

Immediately, Gemma started reasoning about how to formulate a query for CourtListener. That's a good start. It did not try to do the work itself. It delegated. Here's the reasoning it used.

The Duty to Settle in Texas and the Coverage-Uncertainty Conflict

Duty to settle (Texas): Generally, an insurer has a duty to settle a claim within reasonable limits if settlement is advisable to avoid a judgment in excess of the policy limits.

Key Cases

The Stowers doctrine (the "Stowers" duty).
Stowers v. Teague (1977) — the foundational case.
State Farm v. Simmons (2001) — clarifying the duty and the "reasonable person" standard.
Recurring Stowers issues: bad faith, excess judgment, and the "reasonable" settlement.

The Stowers Doctrine: The Three-Prong Test

An insurer's duty to settle arises when:

The insurer has knowledge of facts that would lead a reasonable person to believe that a judgment in excess of the policy limits is likely.
The insurer has the opportunity to settle within policy limits.
The insurer fails to settle.

The Coverage-Uncertainty Problem

By issuing a Reservation of Rights (ROR), the insurer is essentially saying "we might not cover this." Does the Stowers duty exist if the insurer is unsure of coverage?
The duty to defend is broader than the duty to indemnify. Even where coverage is uncertain, the duty to defend often exists if there is a "potential" for coverage.
The conflict: if the insurer refuses to settle because it doubts coverage, but a jury later finds coverage exists and the settlement would have prevented an excess judgment, the insurer is liable for bad faith / the excess judgment.
The dual-interest problem: the insurer's interest in denying coverage versus the insured's interest in settling.

Procedural and Doctrinal Tools

Reservation of Rights (ROR) letters.
Declaratory judgment (DJ) actions — insurers often file these to resolve coverage before trial.
"Duty to defend" vs. "duty to indemnify."
Stowers application when coverage is contested.
Bad faith claims under the Texas Insurance Code and common law.

A Key Distinction

Distinguish the duty to defend (triggered by the possibility of coverage) from the duty to settle (triggered by the risk of an excess judgment). The tension arises because an insurer might use the "uncertain coverage" argument as a shield to avoid the Stowers duty — but Texas law is strict about the insurer's obligation once the risk of an excess judgment becomes apparent.

If you know the area, you can already see the problem. It has zeroed in on the wrong case! The foundational case is not Stowers v. Teague (1977); it is G.A. Stowers Furniture Co. v. American Indemnity Co., 15 S.W.2d 544 (Tex. Comm’n App. 1929). And if we read the referenced Simmons case, we see another problem. It's about first-party insurance, not liability insurance. My prompt made clear I only cared about the latter. But, maybe Gemma just got off to a bad start?

Gemma then came up with a plan. After a little more noodling, Gemma found a highly relevant case (Frank's Casing) and formulated a plan. It still has not realized, however, that it has the wrong "Stowers" case.

The initial search results provide several important leads:

Excess Underwriters at Lloyd's, London v. Frank's Casing Crew & Rental Tools, Inc. is a Texas Supreme Court case (2008) that seems highly relevant to the "duty to settle" and "coverage" intersection. The snippet mentions "In Texas, an insurer that settles a claim against its insured when coverage is disp…" (likely "disputed").
Liberty Mutual Insurance Company v. Mid-Continent Insurance Company (5th Cir. 2005) involves a Texas-law diversity case with "important and determinative questions of Texas law" regarding the duty to settle and coverage. This is a crucial procedural lead (certification to the Texas Supreme Court).
Stowers v. Teague is the foundational case (implied by the "Stowers doctrine" search).

The user wants to know about:

The duty to settle (general).
Special procedural and doctrinal issues when coverage is unclear.

Plan

Deep dive into Excess Underwriters v. Frank's Casing to see how the Texas Supreme Court handles the duty to settle when coverage is disputed.
Investigate the Reservation of Rights (ROR) issue and how it interacts with the Stowers duty.
Look for the conflict-of-interest doctrine in Texas (insurer vs. insured) when coverage is contested.
Examine the procedural use of declaratory judgments to resolve coverage before a Stowers duty matures.

Gemma then continued thinking. You can see, by the way, why smaller models incapable of reasoning and, I believe, without access to actual cases, would be hopeless at this task, which is why pre-2025 reports of LLM performance on legal matters were so negative.

The read_document for Excess Underwriters (ID 2834389) was successful and provides a wealth of information.

The Case

Context: The case involves a dispute over whether an insurer (Excess Underwriters) can seek reimbursement from its insured (Frank's Casing) for a settlement paid to a third party (ARCO) when the insurer believed the claims were not covered.

The core legal issue: Can an insurer recover the amount paid in a settlement from its insured if the insurer settles under a reservation of rights and the claims are later determined not to be covered?

The Stowers connection: The insured (Frank's Casing) used a Stowers demand to force the insurer to settle. The insurer settled to avoid Stowers liability (an excess judgment) but reserved its right to seek reimbursement.

The Holdings

The majority (implied by the dissent) denied the insurer reimbursement.

The dissent (Justice Hecht) argues the majority's rule is wrong because it allows the insured to take unfair advantage of the insurer's Stowers liability. The dissent contends the insurer should be allowed to seek restitution if it settles in good faith to avoid Stowers liability, even if coverage is later denied — and that the majority's rule creates a "windfall" for the insured and incentivizes forcing settlements via Stowers demands even when coverage is questionable.

Key doctrinal tension: the insurer's duty to settle (to avoid Stowers liability) versus the insurer's right to reimbursement for non-covered claims. The majority's rule effectively makes the insurer's settlement unconditional, or at least makes recovery very difficult, if the insurer does not want to risk Stowers liability.

Key Procedural and Doctrinal Issues

The Stowers leverage: An insured can use a Stowers demand to force a settlement and then use that settlement to resolve a coverage dispute in its favor — claiming the insurer waived its rights or that the settlement was unconditional.
Reservation of Rights (ROR): Insurers use ROR to settle while maintaining their coverage defenses. The case discusses whether a settlement under ROR entitles the insurer to reimbursement if the claims are non-covered.
Unjust enrichment / restitution: The central mechanism for the insurer's recovery is restitution. The case weighs the insured's "unjust enrichment" (a settlement paid for a non-covered claim) against the unfairness of the insurer being forced to settle by Stowers liability.
The Matagorda County precedent: The majority uses Matagorda County to argue reimbursement could be unfair to a low-asset insured; the dissent argues this is a flawed application.

Summary of Findings

General duty to settle (Stowers): Insurers have a duty to settle within policy limits when a reasonable person would conclude an excess judgment is likely.
The coverage conflict: When coverage is disputed (often via an ROR), the insurer faces a dilemma — settle and risk being unable to recover from the insured if the claim is non-covered, or refuse to settle and face Stowers liability.
The Stowers leverage / windfall: A major issue is whether an insured can use a Stowers demand to force a settlement and then claim the insurer waived its coverage defenses.

Procedural Mechanisms

Reservation of Rights (ROR): used to settle while preserving coverage defenses.
Declaratory judgments (DJ): insurers may use DJ actions to resolve coverage before a liability claim is resolved, avoiding the Stowers dilemma.

The Texas Supreme Court's Stance (from Excess Underwriters)

The court has limited the insurer's ability to seek reimbursement for non-covered claims settled under a reservation of rights, even when the insurer was forced to settle to avoid Stowers liability. The aim is to prevent "windfalls" for insureds — though critics argue it creates an incentive for insureds to exploit the Stowers doctrine.

Gemma has at least found one of the right cases, but for some reason does not seem to have found the majority opinion. Instead, it seems to be inferring what the majority held by reading the dissent – not a method taught in law school.

After a total of about 2.5 minutes, Gemma released its final answer.

In Texas, the duty of a liability insurer to settle a claim is governed by the "Stowers doctrine," which creates significant procedural and doctrinal tension when the insurer's duty to defend or indemnify is uncertain.

1. The General Duty to Settle (The Stowers Doctrine)

Under Texas law, an insurer has a duty to settle a claim within policy limits if:

The insurer has knowledge of facts that would lead a reasonable person to believe that a judgment in excess of the policy limits is likely;
The insurer has the opportunity to settle within those limits; and
The insurer fails to settle.

If an insurer fails to settle under these circumstances, it may be held liable for the entire amount of the judgment, even if it exceeds the policy limits (often referred to as "Stowers liability" or an "excess judgment").

2. The Conflict: Uncertainty of Coverage

A complex legal conflict arises when an insurer is unsure whether the event giving rise to the lawsuit is actually covered under the policy. This typically results in the insurer issuing a Reservation of Rights (ROR) letter, agreeing to defend the insured while reserving the right to contest coverage later.

The core doctrinal issue is the "coverage vs. settlement" dilemma:

The insurer's risk: If the insurer settles a claim under a reservation of rights to avoid Stowers liability, it risks being unable to recover the settlement amount from the insured if the claims are later determined to be non-covered.
The insured's leverage: An insured can use a Stowers demand to force a settlement. If the insurer settles, the insured may argue that the settlement was "unconditional" or that the insurer waived its coverage defenses, effectively gaining coverage it never paid for.

3. Key Doctrinal and Procedural Issues

A. The "Windfall" vs. "Unfairness" Debate

A major point of contention in Texas jurisprudence (notably in Excess Underwriters at Lloyd's, London v. Frank's Casing Crew & Rental Tools, Inc.) is whether an insurer should be entitled to restitution (reimbursement) for settlements paid on non-covered claims.

The argument for insurers: Denying reimbursement creates a "windfall" for insureds, who can use the threat of Stowers liability to force insurers to settle non-covered claims, effectively obtaining "free" coverage.
The argument for insureds: Forcing insurers to settle under a reservation of rights without a guaranteed right of reimbursement creates uncertainty and may unfairly prejudice the insurer if the insured's assets are insufficient to cover the loss.

B. Procedural Mechanisms: Declaratory Judgments

To mitigate the risk of Stowers liability while coverage is in dispute, insurers often use declaratory judgment (DJ) actions. By filing a DJ action early, the insurer attempts to resolve the coverage question before a liability claim matures, thereby avoiding the coercive pressure of a Stowers demand and the risk of an excess judgment.

C. Reservation of Rights and "Unconditional" Settlements

A significant procedural issue is whether a settlement reached under a reservation of rights can be deemed "unconditional." If an insurer settles to avoid Stowers liability but fails to clearly preserve its rights, the insured may argue that the insurer is estopped from later contesting coverage, making the settlement final and non-reimbursable.

Summary: The Stowers Dilemma in Uncertain Coverage Cases

Issue	Insurer's Position	Insured's Position
Primary goal	Avoid an excess judgment (Stowers) while protecting coverage.	Force a settlement within policy limits to avoid trial.
Risk management	Use a Reservation of Rights (ROR) to preserve defenses.	Argue the settlement was unconditional or that coverage was waived.
Legal remedy	Seek restitution for non-covered claims paid.	Argue unjust enrichment does not apply if the insurer had a duty to settle.
Procedural tool	Declaratory judgment to resolve coverage early.	Stowers demand to force settlement under threat of excess liability.

So, let me say this about the answer. If I had asked the initial question as an exam question (which I would not do) and a student gave me the above answer, I would give it a B- or maybe a B in our grade-inflated world. It's not terrible. But we are a long way from an A.

Calibration

Claude Opus with Connectors

I asked Claude Opus to evaluate the Gemma answer and told it to likewise use CourtListener to help. Here's its summary (after I corrected Claude when it too got something wrong!).

Bottom line. The response is a well-organized overview that correctly maps the architecture — Stowers, reservation of rights, the reimbursement debate, declaratory judgments — but it inverts the governing reimbursement rule and mishandles two of the doctrines the question specifically targets.

The central error: reimbursement backwards

In Texas Ass'n of Counties Risk Management Pool v. Matagorda County, 52 S.W.3d 128 (Tex. 2000), the Court held an insurer settling under a reservation of rights generally has no right to reimbursement for a later-determined non-covered claim, absent a policy provision or the insured's clear and unequivocal consent. The response leans on Excess Underwriters v. Frank's Casing to stage a "windfall vs. unfairness" debate but omits that Texas resolved it against insurers: the Court withdrew its 2005 insurer-friendly opinion and, on rehearing in 2008 (246 S.W.3d 42), refused any exception to Matagorda County. Its summary table even lists "seek restitution for non-covered claims" as an available remedy — the opposite of the rule.

The rule it states is incomplete

The response states a Stowers rule up front, so the problem isn't a missing case name. The problem is that the rule it states — likely excess judgment, opportunity to settle, failure to settle — omits the threshold element that the duty arises only if the claim is within the scope of coverage. That element is the hinge of this entire question: there is no Stowers duty for a non-covered claim, so the insured's "leverage" to force settlement is contingent on coverage existing. The response treats Stowers as freely operative and never supplies the part that the coverage-uncertainty scenario turns on.

Declaratory judgments — mentioned but mishandled

The response does flag DJ actions, but presents them as a clean escape hatch that lets the insurer "resolve the coverage question before a liability claim matures." That skips the actual special procedural problem: in Texas a coverage DJ frequently cannot precede the liability trial, because courts will abate it when the coverage determination depends on facts that overlap the merits of the underlying suit and would prejudice the insured. Treating the DJ as a tidy fix misses precisely the procedural friction the question was asking about.

Smaller gaps

No accurate citation to Stowers itself (15 S.W.2d 544); the eight-corners duty-to-defend rule, conflicts and independent-counsel rights, and In re Farmers (Tex. 2021) all go unmentioned.

Overall: accurate on structure, but wrong on the reimbursement rule, incomplete on the rule it states, and superficial on declaratory judgments — the two doctrines the prompt most directly raises.

Compared to the best frontier models

I've discussed in a prior blog post why using high end frontier models (like the latest from Claude) with connectors such as CourtListener or Midpage to real legal databases can produce good results. I wanted to test that again. I thus gave Claude Opus 4.8 the same prompt that I had given Gemma with a fresh context window. I told it to use CourtListener. The result is a way better response. It's 4,000 words long with six structured sections plus a synthesis "playbook," an open-questions section, and a full table of authorities of 18 cases, including five opinions read in depth for verbatim quotes. It found the right Stowers case. The answer targets the actual question. Coverage uncertainty isn't treated as one bullet; it's the spine — coverage-as-an-element, the settle-or-refuse dilemma, reservation-of-rights conflicts, the reimbursement rule, the contribution-from-insured trap, multiple claimants, and the assignment/adversarial-trial guardrails. It's intellectually honest and contains most of the intricate issues that have long made it one of my favorite areas of law. One citation it couldn't fully verify (In re Farmers' reporter cite) is flagged in the document itself rather than smoothed over. I've read the response and it is basically correct. If I ever saw this on an exam – and I live for that day – it would get a clear A.

Because Claude's answer is too long to realistically put inside this blog entry I asked Claude to turn it into this website. But here's a screen capture if you don't want to click through.

Compared to Lexis Protege

I also compared the local model's answer to that provided by Lexis Protege. I gave it the same prompt. After completely crashing the first time – pretty shocking for a professional product – and producing a research summary, I asked Protege to draft a memo. The result is not great but not terrible. It is still better than the local model, but it's also not good. I've incorporated it in full below.

protege-duty-to-settle

protege-duty-to-settle.docx

19 KB

Protege's work is properly structured as a formal legal memo, and nearly every proposition is anchored to authority such as Chapter 541 of the Texas Insurance Code. It correctly states the Garcia three-part test and the threshold rule that no coverage means no Stowers duty (Stoker; Traver). It has lots of links to click. Its weaknesses, however, are serious, and one is nearly disqualifying as written: the memo conflates first-party and third-party insurance doctrine. The duty to settle – particularly as I wrote the prompt – is a third-party liability concept — it governs how an insurer responds to a claimant's demand against its insured. Yet the memo spends about half of its 13-pages on collateral issues relating to a duty to settle in first party insurance or in largely irrelevant statutory matters. While someone experienced in Texas insurance law might regard the surplusage as merely annoying, for the inexperienced attorney – the one who would actually enter a prompt of the kind I did – it would end up confusing them in a way that could take many billable hours to dispel. Moreover, the memo also misses the question's core — the coverage-uncertainty doctrine of Matagorda County, Frank's Casing, and other cases — substituting generic bad-faith claims-handling for the third-party settlement problem actually posed. If I were grading it, I would probably give it a B with "So what?" or "Not responsive" scrawled across the pages. When calibrated against a largely free model, that's an unacceptable result.

Conclusion

Here is where I wanted to tell you the laptop won. It didn't. The local model produced something shaped like legal analysis — fluent, organized, sure of itself — and underneath it was not good enough to trust: thin on authority, loose on the rules that actually decide cases, and most confident in exactly the places it should have hedged. Protégé was a little better but still planted a serious doctrinal error and wasted time on distracting irrelevancies. The large model running in the cloud and connected to CourtListener was the only one of the three whose results I would even think of putting in front of a client, and then only after some serious checking and editing.

Be wary, too, of promises that an elaborate harness would cure the deficiencies I found with local models. Much of what looks like local-model progress is really progress in scaffolding — retrieval, tool calls, agentic loops wrapped around a small model to make it look larger than it is. Some of that helps. But the model inside the harness is still the model inside the harness, and a clever rig that delivers the wrong cases faster has improved nothing. Could I close the gap by hand-feeding the local model the right statutes, the right opinions, the right framing? But every hour I spend assembling context for a weak model is an hour I did not spend practicing law. The local model was supposed to free me from that sort of dependence. Becoming its research assistant is a strange way to be freed.

All of this is quite unfortunate. Maybe in 2027 I'll be able to write, "At last, local models deliver." Or maybe I or someone else will come up with a better harness. I haven't entirely given up. I'm also poking about whether using a legal database like that at Descrybe would work better than CourtListener. The real sting is that the lawyers who would gain most from a free, private, offline assistant — the solos, the legal-aid offices, the clinics counting every dollar — are the ones the rising token bill might price out first.

Afterwards

I've been trying to simulate what would happen if I indeed had that $50,000 refrigerated box and could run a much larger model locally. I've done this by writing a little app that hands a legal question to exactly the kind of open-weight model you'd run on that box — Gemma, Llama, GPT-OSS, Nemotron — but reaches it through OpenRouter instead of through hardware I don't own. The model doesn't answer from memory. The app gives it live tools to search CourtListener, pull the full text of opinions, and follow citations from modern cases back to their foundational sources, so it researches like a junior associate instead of confabulating. It cites only authority it actually retrieved and read, then assembles a memo with a verified source list and a downloadable library of the underlying cases. I can swap models, watch what each query costs, and compare outputs.

Since these models run through API calls to OpenRouter are some of the same ones I'd be running locally, it's a reasonably faithful preview of the capability — letting me ask whether the legal work would be good enough to justify the box, before buying it. The project is not yet finished but, so far the results from midsize models such as Llama (Meta), and ChatGPT-OSS 120B (OpenAI), and a bigger version of Gemma (Google) have not been encouraging. By constrast, initial results from the 550 billion parameter Nemotron from NVIDIA look quite promising, although it is the kind of model that would probably need more than a $50,000 box to run locally. The law may be complicated enough that it just takes an expensive, big model to get decent results. There's also the possibility of trying a model fusion technique that hands out the same query to multiple tool-provisioned smaller models, retrieves the cases, performs the analysis and then has those multiple analyses reconciled and fused by a smart and potentially larger model. But I haven't even started running an experiment with that technique.

What about a GPU VPS?

One possible halfway house is a GPU VPS. That means renting, by the hour or by the month, a remote computer with a serious graphics processor attached. In plain English: instead of buying the $50,000 refrigerated monster box and keeping it in your office, you borrow someone else’s monster box over the internet.

That helps, but it does not magically solve the problem. It may let a lawyer run much larger open-weight models than will fit on a laptop -- the sort that might actually be able to do a decent job. It may also avoid the full capital cost of local hardware. If the machine is spun up only when needed and shut down when the work is done, it might be cheaper than watching a frontier model burn tokens through a long legal research project. It also gives the lawyer more control over the model and the surrounding research harness.

But the privacy story is no longer “the client’s facts never leave my machine.” They do leave. They go to a server rented from a cloud provider. That may be better than sending them to an AI vendor that trains, logs, or meters everything, but it is not the same as local. Now the lawyer must worry about firewall rules, SSH keys, passwords, updates, logs, backups, disk encryption, vendor terms, and whether a public web interface has accidentally been left open to the world.

For many solos and small firms, that is just too scary. Not because lawyers are foolish, but because this is systems administration, not law practice. The realistic version would require a trusted technical consultant: someone who can set up the server securely, configure the model and legal-research tools, create automatic shutdown and budget controls, disable unnecessary logging, document the system, and be available when CUDA, Python, certificates, or some miserable connector decides to die on a Friday afternoon.

So a GPU VPS softens the dilemma. It does not eliminate it. It trades vendor dependence for infrastructure dependence.