Start with the conclusion: as of October 2026, Meta's hosted Llama API public preview has been retired since July 6, 2026. When people discuss "free Llama quota" today, there are really two main paths, hosted and self-hosted. Hosted splits further into third-party free tiers, cloud platform trial credits, and trial credit at domestic gateways. Hosted is quick to start but has ceilings on rate, data, and stability. Self-hosting does not charge per token under the license, but hardware, power, and operations are yours to carry.

Note: Llama and Meta Llama are trademarks of Meta Platforms, Inc. This article is a technical explanation and does not imply any affiliation, agency, sponsorship, or endorsement by Meta. Model licenses, available versions, and quota policies may change; follow Meta and each provider's pages.

1. Four kinds of "free" are not the same

Path Cost shape Main limits Suitable for
Third-party hosted free tier Daily or monthly volume cap with rate limits; service stops or converts to paid when exceeded Model range, concurrency, context, data terms, no SLA Prototypes, low-concurrency internal tools, evaluation
Cloud platform trial credits New-user credits or time-limited quota under cloud rules Expiry, region, available models, settlement method Short POCs, cloud integration validation
Self-hosted open weights The license does not charge per token; cost shifts to hardware and labor GPU memory, power, operations, license obligations Sensitive data, steady usage, offline or edge
Domestic AI gateway trial Platform trial credit or a small top-up Model, quota, and rate per the platform page Teams needing RMB billing and one multi-model entry point

All four paths may be called "free," but the contract, data boundary, and shutdown risk are completely different. First confirm whether you need short-term validation or long-term production, then talk about price.

2. Meta's own Llama API is retired

On July 6, 2026, Meta shut down the Llama API public preview; existing requests receive a discontinuation notice and a redirect. Meta's guidance was to migrate to third-party providers that support Llama, and its developer platform has since shifted focus toward the Muse family. Muse and Llama are not the same model family; do not conflate them when purchasing or migrating.

The Meta Llama API endpoints, preview keys, and free quotas in older tutorials no longer apply. Llama weights can still be obtained from Meta's download page or the meta-llama repositories on Hugging Face, and downloading requires accepting the community license for that version. That is why a 2026 search for "free Llama API" mostly returns third-party platforms.

3. Hosted free tiers: convenient, with five concentrated limits

  1. Rate and concurrency: requests per minute, tokens per minute, and concurrency caps all apply, and exceeding them usually returns 429. During peak hours, free tiers generally have lower priority than paid tiers.
  2. Model range: free tiers may expose only small models, quantized versions, or older snapshots. Some inference providers retired certain Llama models in 2026, so free lists change quickly.
  3. Context and features: long context, multimodality, tool calling, and structured output may be trimmed. Verify each item during selection.
  4. Data terms: retention and training terms for free tiers often differ from paid tiers and may permit use for service improvement. Do not send customer data, contracts, or code into them directly.
  5. Service assurance: there is usually no SLA, no compensation, and no reliable outage notification. Maintenance, rate changes, and model delisting may come with little notice.

Free tiers are good for validating a need, not for carrying production. Any page claiming "free forever, unlimited calls" should be assumed to change at any time.

4. Self-hosting: the license is free, the resources are not

The Llama community license permits commercial use and does not charge per token, but it does not mean zero cost. The larger the weights, the higher the memory, bandwidth, and operations investment. The following are rough hardware ranges; actual requirements depend on quantization, inference framework, context length, and concurrency:

  • 1B–8B class: 4-bit quantized weights are about 1–6 GB. With KV cache and framework overhead, 8–16 GB of GPU memory is a starting point, suitable for a single card, a laptop, or edge devices.
  • 70B class: 4-bit weights are about 35–45 GB. Long context and high concurrency need more, typically a 48 GB single card or multiple cards.
  • Llama 4 Scout: 109B total parameters and 17B active parameters; 4-bit weights are about 55–60 GB. The nominal long context is demanding on KV cache, so real deployments shrink with concurrency.
  • Llama 4 Maverick: 400B total parameters and 17B active parameters; 4-bit weights are on the order of 200 GB, typically requiring a multi-GPU server.
  • Llama 3.1 405B: BF16 weights are on the order of 800 GB, in multi-machine, multi-GPU territory, and not suitable for a small team starting out.

MoE models activate fewer parameters, so compute demand is relatively lower, but all weights still have to fit in memory. Do not read "17B active" as "17B of GPU memory." The total cost of self-hosting also includes GPU purchase or rental, electricity, racks, network, storage, redundancy, monitoring, upgrades, and security patches. At low utilization, self-hosting is often more expensive than a pay-as-you-go API.

5. Privacy and compliance: self-hosting is not automatic compliance

Running a model on your own servers makes the data boundary more controllable, but logs, monitoring, backups, remote operations, and account permissions can still take data outside. Renting cloud GPUs for self-hosting is the same: the data still enters the cloud provider's infrastructure.

For hosted free tiers, check: whether data is used for service improvement, how long it is retained, whether sub-processors are involved, which region the data lands in, and what the deletion mechanism is. Free-tier data terms often differ from paid tiers, so do not send test data, customer records, or code into them directly. Teams in China also need to assess the path for personal information leaving the country. Where personal information, important data, or regulated industry data is involved, have legal and security teams confirm before deciding between hosting and private deployment. The license's Acceptable Use Policy and trade compliance terms should also be reviewed.

6. How to compare quality: same weights, different results

The same Llama weights can perform differently across platforms because of quantization precision, inference framework, sampling parameters, system prompts, context length, safety filtering, and model snapshot version. A free host may offer only a quantized or older version; the page may say Llama 4, but the actual model ID returned by the API is the reference.

Do not select on leaderboards alone. Prepare 20–50 of your own real tasks with unified prompts and scoring criteria, and compare accuracy, p95 latency, token consumption per task, and manual rework rate. Llama 4 Scout and Maverick lean toward multimodality and long context, while Llama 3.3 70B is better suited to text-heavy tasks. Match by task, not by parameter count.

7. How to choose: by workload, not by price

Use these questions to triage quickly:

  1. Can the data enter a third party? If not, look at private deployment first.
  2. Is traffic steady enough to justify maintaining GPUs? For low, spiky volume, prefer hosting.
  3. Do you need long context, multimodality, or tool calling? Verify that the free tier exposes them first.
  4. Can you tolerate 429s and model delisting? If not, prepare a paid tier or a second provider.
  5. Do you have operations capability? If not, do not self-host.

A common combination is: free hosting for development and low-volume evaluation, paid hosting or a domestic gateway for production, a separately deployed private instance for sensitive modules, and one backup channel kept in reserve.

8. Upgrade path from free to paid

Consider paying when you see these signals: frequent 429s, latency jitter, trimmed context or features, data terms that fail review, a need for SLAs and ticket support, a need for stable model versions, or a need for RMB invoices.

An upgrade does not necessarily mean changing providers. You can move to pay-as-you-go on the same platform, or migrate traffic to cloud hosting, a specialized inference provider, or a domestic AI gateway. The ZSQH AI Gateway provides an OpenAI-compatible interface and can serve as a unified multi-model entry point. Models and prices change with providers and versions; follow the model pricing page, and sign up if you need an account.

curl -sS "https://ai.zhongshangqihang.com/v1/chat/completions" \
  -H "Authorization: Bearer YOUR_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "llama-3.3-70b-instruct",
    "messages": [{"role": "user", "content": "Explain the main limits of free Llama APIs in three sentences"}],
    "max_tokens": 256
  }' | python3 -m json.tool

Model IDs follow the platform's model list. For self-hosting, replace the address with the OpenAI-compatible endpoint exposed by vLLM, SGLang, or llama.cpp. Production code should distinguish 401, 429, and 5xx: check the key on 401, back off and retry on 429, and trigger a backup channel on 5xx.

9. Frequently asked questions

Is there a permanently free Llama API?

As of October 2026, there is no Meta-operated Llama free API that can be treated as a long-term commitment. Third-party platforms may offer free tiers or trial credits, but the quota, rate, and duration are decided by each platform and can change at any time.

Does Meta still offer a Llama API?

Meta's Llama API public preview was retired on July 6, 2026, and the guidance was to migrate to third-party providers that support Llama. Meta's developer platform currently emphasizes the Muse family, which is not the same family as Llama. Whether there is a new Llama hosting plan depends on Meta's pages.

Is self-hosted Llama really free?

The license does not charge per token, but hardware, power, data center space, operations, and redundancy all cost money. When usage is very small or highly variable, self-hosting is usually not worthwhile. It has the advantage when data cannot leave the country, offline operation is required, or utilization is high and stable over the long term.

Is free hosted Llama the same model as the paid version?

Not necessarily. A free tier may expose only smaller models, quantized versions, or older snapshots, and context and features may be trimmed. Rely on the model ID, version number, and model card returned by the API, not the marketing name.

Where should a small team start?

Start with a free hosted tier for one to two weeks of low-volume validation, and run a local baseline with an 8B-class model on a single GPU or laptop to compare quality and latency. Once usage, data needs, and budget are confirmed, choose paid hosting or self-hosting, and keep a backup channel.

Related reading: API Relay Usage Guide, How to Choose a Multi-Model API Platform.