llama-4-scout-17b-16e-instruct

Introduction

Llama 4 Scout is Meta’s open-weight mixture-of-experts model, activating 17 billion of its 109 billion total parameters per forward pass, with native multimodal (text and image) input and multilingual output across 12 languages. Meta advertises an industry-leading 10-million-token maximum context length for Scout, though most third-party hosting providers currently serve it at a more modest ~128K-131K tokens. As an open-weight general-purpose model, it has no built-in legal database, citation checking, or law-firm workflow — it is used either self-hosted, via cloud API providers, or as the engine inside a legal-tech product.

Performance in Legal Research

Llama 4 Scout can reason over legal text and discuss legal concepts capably, but has no live connection to case law or statute databases, so unverified legal research carries real hallucination risk, the same caution that applies to every general-purpose model on this list. Its main research advantage is contextual: an extremely large advertised context window makes it theoretically well suited to holding vast case files or research materials in memory at once, where a hosting provider actually exposes that full window.

Contract Analysis

Scout can extract clauses and flag anomalies in contracts provided in-context, and its multimodal input means it can also read scanned or image-based contract pages directly. As with research, it lacks a legal-specific risk taxonomy and requires attorney review of any extracted findings.

Document Drafting

The model can draft general legal-style text and correspondence from prompts and reference material, but has no legal template library or firm-playbook integration built in; drafting quality and reliability depend on the surrounding application built around the raw model.

Legal Translation

Llama 4 Scout is instruction-tuned for multilingual output across 12 languages, giving it reasonable capability for translating or summarizing foreign-language legal documents, again subject to human verification of legal-specific terminology.

Pricing

As an open-weight model, Llama 4 Scout can be self-hosted at infrastructure cost only, or accessed via API providers at roughly $0.05-$0.10 per million input tokens and $0.10-$0.30 per million output tokens, among the cheapest options for its capability class. There is no vendor subscription fee since Meta does not charge for the model weights themselves.

Verdict

Llama 4 Scout is an inexpensive, flexible, open-weight option with a genuinely enormous advertised context window and native multimodal input, appealing for teams that want to self-host or build custom legal tooling cost-effectively. It has no legal-specific grounding of its own, actual usable context length varies significantly by hosting provider, and it should be treated as infrastructure for a legal application rather than a finished legal research or drafting tool.