LLaMA

Top Builders

Explore the top contributors showcasing the highest number of app submissions within our community.

LLaMA (Large Language Model Meta AI)

LLaMA is a state-of-the-art foundational large language model designed to help researchers advance their work in the subfield of AI. It is available in multiple sizes (7B, 13B, 33B, and 65B parameters) and aims to democratize access to large language models by requiring less computing power and resources for training and deployment. LLaMA is developed by the FAIR team of Meta AI and has been trained on a large set of unlabeled data, making it ideal for fine-tuning for a variety of tasks.

General
Release date	2023
Author	Meta AI FAIR Team
Model sizes	7B, 13B, 33B, 65B parameters
Model Architecture	Transformer
Training data source	CCNet, C4, GitHub, Wikipedia, Books, ArXiv, Stack Exchange
Supported languages	20 languages with Latin and Cyrillic alphabets

Start building with LLaMA

LLaMA provides an opportunity for researchers and developers to study large language models and explore their applications in various domains. To get started with LLaMA, you can access its code through the GitHub repository.

LLaMA Links

Important links about LLaMA in one place:

LLaMA Research Paper - Read the research paper about LLaMA to learn more about the model and its development process.
LLaMA GitHub Repository - Access the LLaMA model, code, and resources on GitHub.

Edit on GitHub

Meta LLaMA AI technology Hackathon projects

Discover innovative solutions crafted with Meta LLaMA AI technology, developed by our community members during our engaging hackathons.

SOAP Copilot: AI Clinical Scribe on AMD

Physicians spend roughly two hours a day on clinical documentation, time pulled directly away from patient care, and a well-documented driver of burnout that costs the US healthcare system an estimated $8.3B a year in lost productivity. SOAP Copilot is a multi-agent AI system that turns a raw doctor-patient conversation into complete, structured clinical documentation in under 30 seconds. Three specialized agents, all built on Llama 3.3 70B Instruct, handle the pipeline in sequence: a SOAP Generator produces a structured Subjective/Objective/Assessment/Plan note, an ICD Coder extracts diagnosis codes with confidence scores, and a Summary Writer rewrites the note as a plain-language patient summary. The entire stack runs on open-source models, so no PHI ever needs to leave the user's own infrastructure. The project was originally prototyped and LoRA fine-tuned on a single AMD MI300X (192GB HBM3) via AMD Developer Cloud, running vLLM on ROCm 7.2. That memory headroom is what makes serving a 70B-parameter model in BF16 precision practical on one chip, something no consumer GPU and few cloud GPUs can do. For this submission, production inference is served through Fireworks AI, which itself runs on AMD Instinct hardware, so the project demonstrates the AMD AI stack end to end: self-managed fine-tuning through to hosted inference. Roadmap items include a distilled specialist model for lower-cost deployment, retrieval-augmented generation over clinical guidelines using the Qdrant vector database already wired into the architecture, and an on-premise, HIPAA-compliant deployment path for health systems that cannot send patient data to any external API.

SmartBandit Router

BanditRoute treats each AI model as an "arm" in a multi-armed bandit problem. Using Thompson Sampling conditioned on task type (code, math, QA, creative), it learns over time which model to send each query to balancing correctness against token cost and latency via a custom reward function, rather than optimizing for accuracy alone. No routing rule was ever hand-written; behaviors like sending math questions to the strongest model and easy QA to a cheap one emerged entirely from feedback. The project is validated two ways: a full simulation (2000 queries, 4 mock model profiles) that shows clean convergence and ~93% cost savings, and a live integration that runs the same bandit logic against real free-tier models via the OpenRouter API proving the approach works beyond simulation, not just in theory.

Thymus

Thymus is a lightweight hybrid token-efficient router designed to maximize accuracy while minimizing token costs in multi‑task LLM pipelines. It dynamically routes user queries across local and remote models on LLM providers. The general architecture of the router is as follows. 1. Classify: a fast regex/keyword pass handles the obvious cases for free; anything ambiguous falls back to the local model for classification. 2. Infer locally: the local model (Qwen2.5-3B-Instruct, GGUF, CPU) answers the query and self-reports a routing preference and confidence. 3. Verify: a task-aware verifier sanity-checks the answer (JSON structure, code execution, entity coverage, self-consistency, depending on task type). 4. Decide: LOCAL (free, done), SHEPHERD (local answer used as context, one cheap cloud pass to double check), or RESET (send straight to Fireworks).

Frugal Router: Cost-Aware Multi-Model AI Agent

Enterprises don't need a frontier model for every request. This agent detects which of the 8 task categories (factual, math reasoning, sentiment, summarization, NER, code debugging, logical/deductive reasoning, code generation) a prompt belongs to, then routes accordingly: code and constraint-satisfaction puzzles go straight to a frontier Fireworks model since small models are unreliable there and a failed cheap attempt costs tokens in addition to the retry; sentiment/NER/factual questions try a bundled 1.5B local model first (zero Fireworks tokens) and fall back only when it isn't confident. The difficulty classifier for math/factual ambiguity is trained on real data — MATH's Level 1-5 labels and RouteLLM's Chatbot Arena win/loss judgments — rather than hand-tuned guesses. Model tiers are built dynamically from ALLOWED_MODELS at container startup since the permitted model list isn't published until launch day.

Cyber Sentinel AI

Cyber Sentinel AI is an AI-powered cybersecurity investigation platform designed to help users quickly analyze potential cyber threats through a collaborative multi-agent workflow. Instead of relying on a single AI model, the platform separates responsibilities across specialized agents for threat analysis, independent verification, and incident response generation. Users can investigate suspicious emails, URLs, prompt injection attempts, and browser extension permissions through a unified interface. Each investigation follows a structured pipeline that combines rule-based evidence collection with AI reasoning to generate actionable recommendations, a dynamic security score, an investigation timeline, and persistent history. The backend is built with FastAPI, containerized using Docker, and designed with a modular architecture that makes it easy to add future investigation modules such as ransomware behavior analysis, live threat intelligence, and ROCm-powered local inference on AMD GPUs. Cyber Sentinel AI demonstrates how multi-agent systems can improve the accuracy, transparency, and usability of cybersecurity investigations while providing a scalable foundation for future security automation.

YouTube Growth Copilot

YouTube Growth Copilot is an AI-powered analytics platform designed to help creators transform raw YouTube Analytics data into practical growth decisions. Instead of simply displaying metrics, the platform combines deterministic data analysis, machine learning, and large language models to explain what creators should do next. The application imports real YouTube Data API and YouTube Analytics API information into a SQLite database, where it detects upload-time patterns, title performance, traffic-source distribution, weak engagement signals, and content trends using Pandas, NumPy, SQL, and custom analytics. A GradientBoostingClassifier predicts whether a newly planned video is likely to perform as Low, Medium, High, or Viral during its first seven days based on upload timing and title features. The structured analytics are then passed to an AI assistant powered by Llama via Groq, which generates upload strategies, content ideas, thumbnail concepts, title recommendations, channel audits, and weak-video diagnoses using structured JSON outputs. The complete application is built with FastAPI, React, SQLite, Docker, and Google OAuth, providing creators with an end-to-end workflow for analyzing their channels and making better publishing decisions based on historical performance instead of guesswork.