Practical tech guides for small businesses in the US and Canada

LlamaCon 2025: What Meta Announced and What Developers Should Check

A MacBook with lines of code on its screen on a busy desk
Photo: Christopher Gower / Unsplash

LlamaCon 2025 gave developers a clearer picture of Meta’s tools around Llama, including a hosted API preview and additional security tooling. The enduring question for a US or Canadian development team is how those tools fit a real application: what can be deployed, under which terms, with what evidence of reliability?

What happened at LlamaCon 2025?

Meta’s April 29, 2025 announcement described its first LlamaCon and introduced Llama API as a limited free preview. The announcement said the service offered access to Llama models, including Llama 4 Scout and Maverick, with Python and TypeScript SDKs and compatibility with the OpenAI SDK.

Meta also announced protection tools including Llama Guard 4, LlamaFirewall and Llama Prompt Guard 2, alongside CyberSecEval 4 and the Llama Defenders Program. These were developer and security resources, not evidence that applications using Llama would automatically be secure.

A computer screen with a bunch of code on it
Photo: Chris Ried / Unsplash

Separate a historical preview from today’s offer

A launch announcement records what was offered at a particular time. It does not establish which accounts can access a service today, what the current rate limits are or whether its pricing and supported models have changed. Read the current documentation linked from Meta’s developer resources before making a deployment decision.

For US and Canadian teams, confirm country eligibility, billing currency, support availability and the handling of submitted data for the exact service you intend to use. Do not transfer statements about one hosted API to an unrelated provider or a self-managed installation.

A practical evaluation plan

Choose a small task with a result you can check. For example, an internal help assistant might need to answer only from a set of approved product documents and identify when those documents do not contain an answer. That is a more useful starting point than asking whether one model is generally “best.”

  1. Define success: record the expected answer, acceptable latency and situations where the application should decline to answer or ask for clarification.
  2. Keep a test set: include ordinary requests, ambiguous questions and deliberate attempts to make the system ignore its instructions.
  3. Compare complete workflows: measure retrieval, model output, tool calls and human review, rather than judging only a short demonstration response.
  4. Check operating costs: include hosting, monitoring, storage and maintenance alongside any model or API charges.
  5. Plan for changes: record model versions and repeat important tests when a provider changes the service or you update dependencies.

What does “open” mean for a project?

Do not assume that a marketing description answers your licensing questions. Review the license and acceptable-use terms for the exact model, release and distribution you select. Hosting arrangements, access to model weights and contractual conditions are separate decisions.

Likewise, a protective model or filtering tool is one part of an application. It should be tested against the actual inputs, tools and data your system exposes. We have not benchmarked the LlamaCon releases and do not claim that they eliminate prompt injection or other security risks.

The security tools Meta highlighted

Alongside the API preview, Meta announced open-source tools for teams building with Llama:

  • Llama Guard 4: a safety classifier that checks prompts and responses, including images, against categories of harmful content.
  • Prompt Guard 2: a smaller model designed to detect prompt injection and jailbreak attempts.
  • LlamaFirewall: a framework that combines guard models and checks to protect AI agents and applications.

These tools can reduce risk, but they’re not a guarantee. Test them on your own data, and keep human review for high-stakes outputs.

Llama license basics

Llama models are released under Meta’s community license, not a standard open-source license. Points teams commonly check:

  • Commercial use is allowed for most companies, but very large services above a monthly active user threshold set in the license need a separate agreement with Meta.
  • The acceptable use policy prohibits certain uses, and you’re responsible for following it in your product.
  • Attribution: the license includes requirements such as displaying “Built with Llama” for some products and naming derived models appropriately.
  • Each version has its own license, so read the one for the exact model you deploy.

Have legal counsel review the license if Llama will power a commercial product.

Where you can run Llama models

OptionExamplesGood for
Major cloud platformsAmazon Bedrock, Microsoft Azure, Google Cloud Vertex AIEnterprise security, billing and US or Canadian data regions
Inference providersGroq, Together AI, Fireworks AIFast, low-cost API access
Your own hardwareOllama, vLLM, LM StudioFull control over data, offline use
Meta’s Llama APIMeta’s hosted serviceDirect access to Meta’s latest tooling, where available

For Canadian organizations with data residency requirements, check whether a provider offers hosting in a Canadian region before sending personal information.

Open models vs closed APIs

  • Open-weight models give you control, predictable costs at scale and the option to fine-tune, but you take on hosting, updates and safety work.
  • Closed APIs such as those from OpenAI, Anthropic and Google often lead on the hardest tasks and handle infrastructure for you, but you depend on the provider’s pricing and policies.

Many teams combine both, using an open model for high-volume, simpler tasks and a closed model for complex reasoning.

Choosing a Llama model size

Meta’s Llama 4 release introduced mixture-of-experts models, such as Scout and Maverick, which activate only part of the model for each request to save computing power. Smaller models in the Llama family run on a single GPU or even a laptop, while the largest need multiple data-center GPUs. A simple rule: start with the smallest model that passes your evaluation, then move up only if accuracy falls short. Bigger models cost more to host and respond more slowly.

A safe deployment checklist

  1. Pin the exact model version and keep a record of its license.
  2. Add input and output filtering, such as Llama Guard or your own rules, before users see responses.
  3. Log prompts and outputs securely, with personal information removed or protected.
  4. Set rate limits and monitoring to catch abuse or unusual costs.
  5. Give users a way to report bad answers, and review those reports regularly.
  6. Document what the system should and shouldn’t be used for, and share it with staff and customers.

How this recap should be used

Use it to understand the April 2025 announcement and identify questions for a present-day evaluation. It is not a promise of current free access, a licensing opinion or a measured comparison against competing systems. The original publication date and URL are preserved.

If your immediate need is an everyday assistant rather than an application platform, our practical prompting guide offers examples of clear task descriptions and verification habits that can also inform your own tests.

Is Llama free?
The model weights are free to download under Meta’s license. You still pay for the computing power to run them, whether on your own hardware or through a provider.

Can I fine-tune Llama on my company’s data?
Yes, the license permits fine-tuning. Keep training data secure, remove personal information you don’t need, and evaluate the tuned model before deployment.

Does Meta see data when I run Llama myself?
No. When you run the model on your own hardware or a third-party provider, Meta doesn’t receive your prompts. Your provider’s terms decide how data is handled.

Will there be another LlamaCon?
Check Meta’s AI blog and developer pages for current events and announcements rather than relying on this 2025 recap.

Related guides