TLM · Law v0.3 · 1.7B

FR-Lex-1.7B

A Thin Language Model for law. Runs standalone, or as a Lens in front of a frontier model — scoping jurisdiction, citations, and routing before the expensive call.

Parameters
1.7B
Quantization
4-bit
Context
32K
Runtime
MLX
License
Apache-2.0
The Lens Pattern

FR-Lex as a pre-filter

FR-Lex is designed for two deployment modes: fully standalone on-device, or as a routing layer that scopes and conditions queries before they reach a frontier model.

User query
Legal question
The Lens
FR-Lex 1.7B
LOCAL · ON-DEVICE
If needed
Frontier LLM
SCOPED + ROUTED

FR-Lex handles jurisdiction scoping, citation extraction, and query classification — resolving most queries locally, and routing only what needs the frontier model's power.

Capabilities

Built for legal reasoning

FR-Lex covers the core reasoning tasks in legal work — jurisdiction, citation, classification, and analysis — without sending client data to a cloud.

001

Jurisdiction Scoping

Identifies applicable jurisdiction from query context — federal, state, civil, common law — and routes accordingly before any expensive model call.

002

Citation Extraction

Extracts and normalises legal citations from unstructured text: case law, statutes, regulations, and secondary sources in major citation formats.

003

Case Classification

Classifies matters by practice area, urgency, and complexity — enabling paralegal routing, docket prioritisation, and matter-opening automation.

004

Contract Analysis

Identifies clauses, flags unusual terms, and extracts key obligations and dates from standard contract structures across common law and civil law templates.

005

Statutory Interpretation

Applies textualist and purposivist lenses to statutory text, surfacing ambiguities and conflicting interpretations with traceable reasoning chains.

006

Privileged Data, Local

Runs entirely on-device. No client communications, case files, or privileged matter data ever leave the machine — compliant by design.

Architecture

How FR-Lex works

Distilled from frontier legal reasoning into a model that fits on a MacBook. On-device inference for privileged data.

01

Legal Corpus

Trained on a proprietary corpus of case law, statutes, regulations, contracts, briefs, and legal commentary — spanning common law and civil law jurisdictions — curated and reviewed by legal domain specialists.

02

Frontier Distillation

Frontier models with strong legal reasoning serve as teachers. FR-Lex learns the underlying judgment — jurisdiction identification, citation normalisation, clause detection — not surface pattern matching.

03

LoRA Fine-tuning via MLX

Efficient parameter fine-tuning on Apple Silicon via MLX, producing a 4-bit quantized model that runs at inference speeds practical for live legal workflows.

04

Standalone or Lens Deployment

Deploy FR-Lex as a fully independent on-device model, or wire it as a pre-filter in front of a frontier model API. In Lens mode, it conditions the downstream prompt with scoped jurisdiction, extracted citations, and a classified query type — cutting frontier model costs by routing resolved queries locally.

Benchmarks

Performance against larger baselines

Evaluated on legal domain reasoning tasks. FR-Lex outperforms general-purpose models significantly larger on its target domain.

Task FR-Lex 1.7B GPT-4o-mini Llama-3.1 8B
Jurisdiction identification 93.2% 85.7% 76.4%
Citation extraction (F1) 91.8 83.1 71.6
Practice area classification 90.4% 82.9% 73.8%
Contract clause identification 88.6% 80.3% 69.5%
Statutory ambiguity detection 84.1% 77.2% 64.9%
Hardware Requirements

Runs on what you already have

Apple Silicon

Recommended · MLX runtime
ChipM1 or later
RAM8 GB unified (16 GB recommended)
Storage~1.1 GB (4-bit)
Tokens/sec~40 tok/s on M2 Pro

CUDA GPU

Supported · llama.cpp / GGUF
GPU VRAM4 GB minimum
RAM8 GB system RAM
Storage~1.1 GB (Q4_K_M)
Tokens/sec~58 tok/s on RTX 3080

Deploy FR-Lex
in your legal stack.

Download the weights, run it on-device, or wire it as a Lens in front of your frontier model to cut costs and keep client data private.