Free models on OpenRouter: the full free-tier list and how to use them

OpenRouter has a killer feature few people know about: dozens of models are available completely free. They carry the :free tag — requests to them don't deduct a cent from your balance.

This article covers what these models are, their limits, and how to connect them from Russia without a VPN or foreign card.

What free-tier models are

Free-tier models are regular models whose cost is subsidized by providers themselves — for infrastructure load, testing, and feedback. You pay zero for tokens; the model is monetized differently.

The naming format is always the same — a :free suffix:

deepseek/deepseek-r1:free
meta-llama/llama-3.3-70b-instruct:free
qwen/qwen-2.5-72b-instruct:free

Understand the free-tier limitations:

Limitation	Typical value
Speed	Slower than paid versions — lower queue priority
Availability	Can go down temporarily under load
Rate limits	Usually 10–20 requests per minute
Context	Often reduced versus the paid version

Perfect for experiments, prototypes, pet projects, and non-critical tasks. For production, take the paid version of the same model.

Which free models exist

The free-tier catalog changes regularly — new ones appear, old ones leave. The current list is always visible in the OpenRouter catalog. By category, you'll usually find:

Reasoning models DeepSeek R1 — an open-source reasoning model that thinks step by step. The free version is a great way to try real chain-of-thought at zero cost.

General-purpose chat models Llama 3.3 70B, Qwen 2.5 72B — strong all-rounders for dialogue, writing, and light coding. Near-GPT-4-class quality, but free.

Specialized models Free versions of code models, translators, and compact classification models appear periodically.

How to connect free models from Russia

Same problem as with all of OpenRouter: signup and payment require a foreign card. The solution is a LiteAI key.

Step 1. Top up on the OpenRouter page — from $25 (2 340 ₽) with a Russian card or SBP. Even if you plan to use only free models, the minimum package is needed to activate the key.

Step 2. Receive your sk-bf-... key by email.

Step 3. A request to a free model looks like this:

curl https://api.liteai.tech/v1/chat/completions \
  -H "Authorization: Bearer sk-bf-..." \
  -H "Content-Type: application/json" \
  -d '{
    "model": "deepseek/deepseek-r1:free",
    "messages": [{"role": "user", "content": "Explain quantum entanglement"}]
  }'

Your balance isn't touched at all. When you want paid flagships — change only the model name; the key stays the same.

Strategy: free + paid in one project

The optimal money-saving scheme with one key:

• Drafts, experiments, learning → free-tier models (deepseek-r1:free, llama-3.3-70b:free)
• Production code and critical tasks → Claude Sonnet or GPT-5.x
• Classification, routing, simple scripts → cheap mini models

That way your paid budget is spent only where quality actually matters, while 80% of experiments cost nothing.

FAQ

Are free models really free? Yes, free-tier tokens cost nothing. You only pay for paid models.

How slow are they? Noticeably slower than paid ones: providers give them lower queue priority. Annoying for interactive chat, fine for batches and experiments.

Can a free model disappear? Yes, the free-tier catalog is dynamic. A provider can close free access any time — never build production around one specific free model.

Do I need a VPN? No. The LiteAI key works directly, payment runs through Russian YooKassa.

One account for everything, or separate? Separate: LiteAI's Claude/GPT token plans and the OpenRouter balance are different products with different keys. Emails must differ. More in OpenRouter from Russia.

Summary

OpenRouter's free tier lets you work with serious open-source models (including DeepSeek R1 reasoning) at zero cost. From Russia, access opens through a LiteAI key: a one-time top-up from $25 with a Russian card, then free models never touch the balance.

Activate access at the top-up page.