Law

Three things make legal text its own problem. The documents are long, so the window has to hold a whole contract or a stack of filings at once. The reading has to be exact: a missed clause or a flipped qualifier changes the answer. And the privacy stakes are high enough that “send it to an API” is off the table before the first prompt. A local model on Conifer covers all three. Only the first two turn on which model you load.

Fit the window to the document#

The context window is the hard constraint. A model reasons over what fits in its window and nothing more, so the moment a brief spills past that limit you are summarizing or chunking instead of reading. For legal work the model’s native context is the number that matters, ahead of a point or two of benchmark intelligence.

Sizes, measured decode speeds, and intelligence scores are on the model ledger. A Gemma-family model carries its own quirks; see The Gemma family before relying on it.

The sparse Qwen 3 30B-A3B-2507 is the standout for this work. It has a 256K native window, and because only a few experts fire per token it reads a long document at roughly the speed of a small model. The catch is memory. A sparse model still loads every expert, so plan for the full weights in RAM. The memory budget, and why a long window costs memory through the KV cache, is in Context & memory.

Careful reading over raw intelligence#

A bigger model is not a better reader. Legal work rewards a model that follows instructions precisely, holds the whole document in view, and does not paraphrase a clause into something subtly different. For questions that turn on a chain of conditions (an indemnity that depends on a notice that depends on a deadline) a reasoning model that works step by step catches what a fast first-pass answer misses. You pay for it in extra tokens before the reply lands.

Match the tuning to the question. Most drafting, summarizing, and clause-by-clause review is an instruct job. Save the reasoning models for the questions where the logic itself is the work.

Citation discipline#

The failure mode that matters here is a confident citation to authority that does not exist. No local model fixes that on its own, and a small one is more prone to it. The fix is procedural, not a model setting. Keep the source in front of the model and make it point back to that text instead of to its training memory.

• Ground every claim in the document. Ask for the quoted passage and its location alongside each answer, so a wrong reading is one you can catch by checking the quote.
• Retrieve, don’t recall. For anything beyond the open document (a corpus of contracts, a deposition set) pull the relevant text into the window through retrieval instead of trusting the model’s memory. The patterns are in Long context & RAG.
• Verify, always. Treat any case name, statute, or date the model produces as a lead to check, never as a finding.

Why local is the point here#

Privileged material, sealed filings, and client data carry duties a cloud API turns into a liability: the moment a document leaves your hardware, you inherit someone else’s retention, logging, and breach posture. Conifer runs the model entirely on your machine. No request reaches a server, no copy lands in a vendor’s logs, and there is no cloud to opt out of because none was ever in the path. The detail of that boundary is in The local-first guarantee.

The same boundary holds when a model reads your files directly. Conifer’s agents are deny-by-default: a model reaches a folder or an account only through a grant you set, and nothing it reads leaves the machine. That is the right shape for a document set you cannot put online. See the grant model. And a contract can carry instructions aimed at the model, so tool output is held as untrusted.