Legal Analysis

1. Install LM Studio (local-first, no data ever leaves) → download Llama 3.3 70B Q4_K_M (40 GB) or Qwen 2.5 32B Q6_K (24 GB).
2. For contract review: load a contract PDF into a local RAG pipeline. AnythingLLM + LM Studio — upload the contract, ask: "Identify any unusual termination clauses, limitation of liability exceeding 12 months of fees, and missing force majeure provisions."
3. First analysis in 15-30 seconds. All data stays on-device — essential for attorney-client privilege.
4. For case law research: index a corpus of case PDFs → semantic search over case law → LLM synthesizes relevant precedents.
5. For e-discovery: embed millions of documents → retrieve relevant ones based on natural language queries → LLM reviews the top matches. Reduces document review time by 70-90%.
6. Critical: Legal AI MUST run locally. Sending client documents to cloud AI services (ChatGPT, Claude API) waives attorney-client privilege in most jurisdictions. On-prem/local deployment is not a preference — it's a professional obligation for attorneys.
7. For specialized legal models: SaulLM, LexiLaw, and LLaMA-2-legal fine-tunes exist but general-purpose 70B models often outperform them on reasoning tasks.

$300-400 genuinely cannot do reliable legal analysis. Legal work requires: (a) 70B+ models for accurate contract interpretation (hallucinating a contract clause = malpractice), (b) long context (32K+ tokens) for full contracts, (c) strict data privacy. A 7B-14B model on 12 GB VRAM ($400 build) will hallucinate legal precedents, miss subtle contractual language, and confuse jurisdictional differences. For legal AI on a budget: used RTX 3090 24 GB + 64 GB RAM ($1,500 total) is the minimum viable setup for contract review. Below that, the risk of AI-generated legal errors exceeds the benefit. $400 buys you document classification and basic keyword extraction, not legal analysis. Be honest about this limitation.

Dual RTX 3090 48 GB total ($1,600, see /hardware/rtx-3090). Runs Llama 3.3 70B Q5_K_M (48 GB) — professional-grade contract analysis, case law reasoning, and regulatory interpretation. For a small-to-medium law firm (5-20 attorneys): this handles e-discovery (1M+ documents), contract review (50+ contracts/day), and legal research. Pair with Ryzen 7 7700X + 64 GB DDR5 + 4TB NVMe (legal document archives are massive). Total: ~$2,500-3,500. For AmLaw 100 firms: 4-8 GPU servers with 8× RTX 3090/4090 for firm-wide deployment. The ROI is dramatic — 70% reduction in first-pass document review pays for the hardware in one case.

The mistake: Uploading client contracts to ChatGPT/Claude/Gemini "for a quick review" because local setup "takes too long." Why it fails: This isn't a quality problem — it's a legal ethics violation. Uploading client documents to a third-party AI service: (1) waives attorney-client privilege (the AI provider's ToS typically claim rights to process/analyze uploaded data), (2) violates data protection regulations (GDPR, CCPA, HIPAA if medical-legal), (3) creates discoverability — the uploaded data is now held by a third party and potentially subpoenaable, (4) violates most state bar ethics opinions on technology competence. Multiple state bars have issued formal opinions on this — the guidance is unanimous: don't upload client data to consumer AI services. The fix: Use local-only AI. LM Studio + AnythingLLM + Llama 3.3 70B provides contract review quality approaching GPT-4 with zero data leakage. If your firm can't afford ~$3,000 for a local AI server, use manual review until you can — the malpractice liability from a ChatGPT data breach far exceeds $3,000. Never put client data into cloud AI. Ever.