Closed labs (OpenAI → xAI) sell access. Meta is the center of gravity for downloadable frontier-class weights: the Llama family, plus vision Segment Anything (SAM), plus the consumer Meta AI assistant in WhatsApp/Instagram/Facebook/Ray-Ban-class surfaces. This lecture is the vendor card for engineers choosing self-host vs Meta AI vs a Llama inference host—not a new transformer derivation.
Vol. 18.4 hardware tiers matter here: Llama on Tier 1 (hosted API) vs Tier 2+ local GPUs. Module 22.5 (Hugging Face, Groq, Together, etc.) is where many teams actually serve Llama. Meta AI is a product; Llama is a weight family; SAM is a segmentation model. Do not collapse the three.
Learning Objectives
By the end of this lesson, students should be able to:
- Separate Llama (open weights + license), SAM (vision segmentation), and Meta AI (consumer assistant).
- Explain that Llama is not automatically OSI “open source”—read the current Meta license and acceptable use.
- Fill a qualitative Pricing / Strengths / Weaknesses card vs closed APIs.
- Sketch a self-host or OpenAI-compatible host call for an instruct Llama checkpoint.
- Know when Meta AI (WhatsApp/IG) is irrelevant to a B2B backend.
- Cross-link Vol. 11.5 Llama, Vol. 7 SAM, and Vol. 18 hardware/SDK patterns without inventing unpublished Llama internals.
Meta in this catalog means three public AI lines. Llama is Meta’s decoder-only LLM family released primarily as open weights (base + instruct, multiple sizes) under Meta’s license terms. Segment Anything (SAM / SAM 2) is Meta’s promptable image/video segmentation family—vision, not an LLM. Meta AI is Meta’s consumer assistant product embedded in Meta apps, typically powered by Llama-class models behind a closed product UX. You can self-host Llama; you cannot download “Meta AI the app.”
Three Surfaces, Three Jobs
| Surface | What you operate | What you do not get |
|---|---|---|
| Llama weights | Checkpoints on HF / llama.com; vLLM, TGI, Ollama, cloud hosts | Meta’s consumer UX, ads graph, or a guaranteed SLA from Meta |
| Llama API / partners | Meta and third parties also host Llama as APIs | Freedom from someone’s token or GPU bill |
| SAM / SAM 2 | Masks from points, boxes, or text-ish prompts; video in SAM 2 | A chatbot or a diffusion generator |
| Meta AI | End-user assistant in Meta apps (and some devices) | A multi-tenant B2B API you control, by default |
Pricing / Strengths / Weaknesses
| Axis | Pricing (qualitative) | Strengths | Weaknesses |
|---|---|---|---|
| Llama self-host | Weights: no per-token fee to Meta. You pay GPUs, electricity, eng time, or a host (Together, Fireworks, Groq, Bedrock, Azure, Vertex, …). Large Llama SKUs need serious VRAM (Vol. 18.4). | Control, fine-tunes (LoRA/PEFT), data stays on your metal if you operate it; huge ecosystem; default open baseline for many teams. | Ops burden; you own latency/safety/eval; license + acceptable-use still constrain some commercial cases; not “free intelligence.” |
| Hosted Llama APIs | Token- or time-metered by the host, usually below flagship closed GPT/Claude list prices for similar size—verify live quotes, do not invent numbers. | Open-weight benefits without buying H100s on day one (Tier 1). | You still trust a host; SKU names and quantization variants multiply. |
| SAM | Open weights: GPU/CPU cost. Cloud vision APIs that wrap SAM-like models are usage-metered. | Best-in-class promptable segmentation starting point; SAM 2 adds video. | Not an LLM provider substitute; integration is CV engineering (Vol. 7), not chat. |
| Meta AI (consumer) | Bundled into Meta apps / device stories—not your production invoice. | Distribution: billions of consumer touchpoints; useful UX reference. | Wrong procurement target for a bank’s backend; data/ToS are Meta’s consumer policies. |
Pick Llama when
- Privacy, fine-tunes, or cost at bulk volume favor self-host
- You want an open baseline everyone can reproduce
- Hybrid: Llama for RAG bulk, closed flagship for hard reasoning
Pick SAM when
- Product needs masks, not paragraphs
- Interactive annotation, editing, or video object tracking
- You already have a CV pipeline (Vol. 7)
Skip Meta AI API dreams when
- You need B2B contracts, VPC, and your own eval
- WhatsApp chat is not the product surface
- You actually needed Llama weights all along
Why Meta dominates open LLM RFPs
- Size ladder + instruct variants + community fine-tunes
- Works with HF, vLLM, Ollama, and Module 22.5 hosts
- SAM is a separate world-class vision gift
Why Meta is not “fully open”
- License/acceptable use ≠ blanket OSI
- Training data and flagship recipe details remain limited
- Meta AI consumer stack is closed product UX
Practical API Sketch
Two honest paths: OpenAI-compatible host (Tier 1) or local runtime. Teaching stand-ins only.
Related Lectures
| Lecture | Role |
|---|---|
| xAI (Grok) | Previous: last closed-first lab |
| Llama family / Open-source models | Weight lineage vs this vendor card |
| SAM (Vol. 7) | Segmentation model textbook |
| Tier 1 / later 18.4 GPU tiers | Host vs self-host sizing |
| Hugging Face (22.5) | Where Llama checkpoints actually live |
| Mistral AI | Next European open+API lab |
Llama = Meta AI = SAM. Three different products. Second: Llama license is always OSI MIT. Third: self-host is free (GPUs are not). Fourth: Meta AI in WhatsApp is your HIPAA backend. Fifth: SAM generates photoreal images like DALL·E. Sixth: open weights remove the need for Vol. 19 eval and Vol. 20 privacy (your logs still exist).
Knowledge Check
- Short Answer: Name Meta’s three AI surfaces in this lecture. Answer: Llama, Segment Anything (SAM), and Meta AI.
- True/False: Llama weights are always identical to OSI “open source” with no license constraints. Answer: False.
- Multiple Choice: SAM is primarily: (a) promptable segmentation, (b) a coding LLM, (c) a video generator like Sora. Answer: (a).
- Short Answer: What does self-hosting Llama actually cost? Answer: Your GPUs/ops or a third-party host—not a Meta per-token fee (unless you use a hosted Llama API).
- True/False: Meta AI in WhatsApp is the default B2B Llama API. Answer: False.
- Multiple Choice: A hybrid pattern taught here is: (a) Llama for private bulk + closed flagship for hard reasoning, (b) SAM instead of all LLMs, (c) skip eval. Answer: (a).
- Short Answer: Which Vol. 11.5 lecture is the Llama lineage home? Answer: Llama family.
- True/False: This page invents Llama token dollar prices. Answer: False.
- Multiple Choice: Where do many teams download Llama checkpoints? (a) Hugging Face / Meta release channels, (b) ChatGPT Plus, (c) NotebookLM. Answer: (a).
- Short Answer: Which provider lecture follows Meta? Answer: Mistral AI.
Key Takeaways
- Meta’s catalog: Llama (open-weight LLMs), SAM (segmentation), Meta AI (consumer assistant).
- Pricing posture: pay GPUs/hosts, not Meta chat seats; licenses still apply.
- Strength: ecosystem + control. Weakness: ops, license nuance, Meta AI ≠ your API.
- Cross-link Vol. 11.5 Llama, Vol. 7 SAM, Vol. 18 tiers, Vol. 22.5 hosts.
- Next: Mistral AI.
Lab: Students pick one Vol. 21 product and decide Llama self-host vs hosted Llama vs closed GPT. Write license+AUP check, VRAM tier (18.4), and a SAM yes/no (only if the product needs masks). Optional: Ollama or a host chat spike.
Whiteboard: Closed labs | Meta (three boxes) | next open labs (Mistral, DeepSeek, Qwen, Phi).
Recap: Meta opens the open-weight half of 22.1 with Llama, SAM, and Meta AI. Next is Mistral AI.