Meta Llama¶
Catalog ID: provider-meta
Scope¶
- Meta-owned Llama model artifacts, official prompt-format documents, and official cookbook methods.
- Deployment notes that preserve the exact Llama model revision, chat template, tokenizer, and runtime boundary.
This is a model-owner guide, not a guarantee about every host serving a Llama-branded artifact. The cited prompt-format page is specifically tied to Llama 3.3.
Do not use this guide for¶
- Community model forks, third-party hosted endpoints, or serving wrappers not documented by Meta.
- Claims that one Llama prompt format or tool convention applies to every Llama release.
License and acceptable-use decisions must be checked for the exact release. This guide does not interpret those terms or describe current hardware requirements.
Stable guidance¶
- Pin the model artifact and tokenizer together, then render messages with the chat template shipped for that revision.
- Store raw role content separately from rendered control tokens so a template migration can be reviewed.
- Re-run behavior and safety cases after quantization, adapter, fine-tune, or runtime changes.
For an open-weight deployment, a prompt is not only prose. The tokenizer and chat template turn roles into the control-token sequence the model actually receives.
Provider-specific prompt behavior¶
- Use the Llama model repository and model-specific prompt-format file as the authority for roles and special tokens.
- Treat Llama 3.3 tool-format examples as version-bound rather than hand-writing those tokens for a different Llama artifact.
- Use the official cookbook as implementation evidence while preserving the license and configuration assumptions of each recipe.
The official Llama 3.3 prompt-format document describes headers, end markers, and tool-call conventions for that version. Keeping messages in structured role form and letting the matching template render them reduces accidental token drift.
Relevant platform features¶
Meta publishes model material and examples, but a self-hosted application supplies the runtime, tool parser, authorization, safety controls, and observability. A model-emitted tool structure is data until the host validates it. The same model weights can also behave differently after a template, quantization, adapter, or decoding change.
The official Llama Cookbook contains practical recipes and integrations. A recipe can support an implementation choice, but it does not turn an untested combination of hardware, server, model revision, and quantization into a verified configuration.
Minimal provider-aware example¶
Keep the fictional Alpine Outfitters inventory request as role messages before rendering:
system:
Answer inventory questions only from the supplied warehouse snapshot.
If an item or timestamp is absent, return "unknown"; do not estimate stock.
user:
Warehouse snapshot: SKU MT-204 has 7 units, captured 2026-07-30T18:00Z.
Question: Can we promise 10 units of MT-204 for pickup tomorrow?
The matching Llama 3.3 tokenizer/template—not a manually copied token string—would serialize these messages. The expected policy decision is not a live-model result.
Production-oriented example¶
An inventory assistant stores a deployment manifest containing the exact model artifact and
revision, tokenizer revision, chat-template hash, runtime, quantization, adapter, decoding values,
and tool parser. It can request a read-only get_inventory_snapshot tool, but the host resolves
the warehouse from authenticated staff scope, validates the SKU, and returns quantity, capture
time, and provenance.
Tests cover missing SKUs, stale snapshots, a tool-like string inside user content, malformed model tool output, multilingual item names, and attempted warehouse switching. Any manifest change creates a new evaluation run. An application may preserve raw messages to compare old and new template rendering before production migration.
Evaluation and portability checks¶
- A reproducibility record names the exact Llama artifact, revision, tokenizer, template, runtime, and decoding settings.
- The example does not manually transplant Llama 3.3 control tokens into another model family or release.
- Any tool request is parsed and authorized by the serving application before an inventory system is queried.
- Compare both rendered token sequences and task outcomes when changing templates; matching prose alone is not evidence of portability.
Fast-stale claims¶
| Area | Why it can change | Source | Verified |
|---|---|---|---|
| Available Llama artifacts and model-card guidance | The official model repository changes as Meta publishes revisions and accompanying documentation. | official-llama-docs |
2026-07-31 |
| Llama 3.3 role, tool, and special-token format | The cited prompt format belongs to one model version and does not establish future formats. | official-llama-prompt-format |
2026-07-31 |
| Llama Cookbook recipes and runtime integrations | Examples track changing libraries, model releases, and deployment techniques. | repo-meta-llama-cookbook |
2026-07-31 |
Official sources¶
official-llama-docs— Llama model repositoryrepo-meta-llama-cookbook— Llama Cookbookofficial-llama-prompt-format— Llama 3.3 prompt format
Known limitations and unverified areas¶
- The inventory example has not been rendered with an actual Llama 3.3 tokenizer in this repository.
- No controlled experiment measured drift caused by quantization or a different inference runtime.
Last verified¶
2026-07-31