Free Access to LLMs

(Appendix)

Intelligent Agents — Module 2

A.Y. 2026/2027

Giovanni Ciatto

Compiled on: 2026-10-22 — printable version

back to ToC

Before you start

• This is a practical reference on how to use LLMs without paying, for study and research
  • for Master’s and PhD students, researchers, and teachers
• All figures are a snapshot as of 30 September 2026
  • quotas, programs, and even models change every few months
  • always check the linked page before relying on an offer
• “Free” here means no money from your pocket: someone else pays (a provider, a cloud, a grant, your lab), or you pay in hardware

What “free” can mean

1. Standing free tier: a recurring zero-price quota (per minute / day / month), no application needed
2. Trial: a one-off balance that expires (e.g. $5 for 30 days)
   • a trial is not a free tier: don’t build a course or a paper on top of it
3. Student credits: cloud credits for enrolled students, renewable while eligible
4. Research grants: credits awarded upon application, for a specific project
5. Local open weights: the API bill is zero, but you pay in hardware and electricity

Which route for me?

Free APIs you can use today

No application: just sign up, get a key, and stay within the quota

One client, many free providers

• All the providers above expose an OpenAI-compatible API: only base_url, key, and model name change

# provider -> (OpenAI-compatible base URL, env var holding the API key)
PROVIDERS = {
    "gemini": ("https://generativelanguage.googleapis.com/v1beta/openai/", "GEMINI_API_KEY"),
    "groq": ("https://api.groq.com/openai/v1", "GROQ_API_KEY"),
    "cohere": ("https://api.cohere.ai/compatibility/v1", "COHERE_API_KEY"),
    "openrouter": ("https://openrouter.ai/api/v1", "OPENROUTER_API_KEY"),
    "mistral": ("https://api.mistral.ai/v1", "MISTRAL_API_KEY"),
    "ollama": ("http://localhost:11434/v1", None),  # local, no key needed
    "vllm": ("http://localhost:8000/v1", None),  # local, no key needed
}

poetry run python -m snippets -l free-access -e 1 groq                               # list the models available now
poetry run python -m snippets -l free-access -e 1 groq openai/gpt-oss-20b "Hello!"   # one chat request

• PROVIDERS: each provider’s endpoint, and the environment variable holding its key (e.g. GROQ_API_KEY, asked interactively if missing)
  • ollama, vllm: open models served locally (on your machine, see below), hence no key
• Arguments: PROVIDER (required), then optionally MODEL and PROMPT (without them, the script just lists the models)
• Full script here (cf. how to run snippets); model IDs are listed live, not hard-coded, since they go stale quickly
• The same trick works with the LLM-as-a-Service scripts: just set OPENAI_BASE_URL, OPENAI_API_KEY, and OPENAI_MODEL (the default, nvidia/nemotron-3-super-120b-a12b:free, exists only on OpenRouter: pick an ID from the listing above), e.g.
  • Ollama: OPENAI_BASE_URL=http://localhost:11434/v1 OPENAI_API_KEY=unused OPENAI_MODEL=gpt-oss:20b
  • Groq: OPENAI_BASE_URL=https://api.groq.com/openai/v1 OPENAI_API_KEY=... OPENAI_MODEL=openai/gpt-oss-20b

Don’t count on these

Students: cloud credits

• Cloud credits can pay for managed model APIs in that cloud (Azure OpenAI, Bedrock, Vertex AI)
• Watch out for the moment credits run out: set a budget alert, or you will be billed

Researchers: grants

• Apply early: approval takes weeks to months

Startups (in brief)

• Since May 2026, AWS Activate credits can pay for third-party models on Bedrock (Anthropic, Mistral, Meta, …) (source)
• “Up to” is the maximum, not what everybody gets

Local models: the only durable free API

• Open-weight models + an inference server = an API that nobody can revoke or re-price
  • gpt-oss-20b (Apache 2.0): about 16 GB of memory, a good laptop / workstation baseline
  • gpt-oss-120b: about 80 GB, one large GPU
  • others: Gemma 4, Mistral Small 4, DeepSeek-R1 distills (check each license, cf. the checklist in the governance lecture)
• Serve them via Ollama (easy) or vLLM (fast, multi-user): both expose an OpenAI-compatible API
  • i.e. poetry run python -m snippets -l free-access -e 1 ollama works out of the box
• Quantization (4-bit) cuts memory by ~3× w.r.t. 16-bit (Q4_K_M ≈ 4.9 bits per weight, cf. the table in LLM-as-a-Service), usually with small quality loss, but re-evaluate on your task
  • see the LLM-as-a-Service lecture for requirements, quantization, and model zoos
• Pinned weights (exact revision + quantization + engine version) are also the best option for reproducibility

Labs and courses: a gateway, not a shared key

1. Create provider accounts under the institution, and keep the real keys only on the gateway
2. Run LiteLLM and give each user (or project) a virtual key with a hard budget and rate limit
3. Keep an allow-list of models; record provider + model ID of each run
4. Clients just point OPENAI_BASE_URL to the gateway: same code as above

Don’ts

These look free, but they break Terms of Service, leak data, or ruin reproducibility:

• Sharing or trading API keys: no attribution, one leak burns everybody’s budget
  • OpenAI recommends one key per person; Mistral forbids transferring keys
• Multiple accounts / cycling trials to multiply quotas: leads to suspension
  • OpenRouter explicitly says extra accounts do not raise limits
• Wrapping consumer chat UIs (ChatGPT, Claude.ai) as an API: forbidden by OpenAI and Anthropic terms, and breaks at every UI change
• VPNs to bypass regional or eligibility restrictions
• Unknown free proxies: the operator sees your prompts, outputs, and keys
• Distilling a free proprietary API into your own model: OpenAI’s terms forbid using outputs to build competing models
  • with open teachers (e.g. DeepSeek-R1), check both teacher and student-base licenses

Privacy and reproducibility

Lecture is Over

Compiled on: 2026-10-22 — printable version

back to ToC