ModelHubby

← All guides · provider record →

Cerebras Inference — free tier setup

What you get free

Ultra-fast inference (~2,600 tok/s per seed repo) on a small set of open models, no credit card. Catalog quota: "Official rate-limits page (Free Trial tier, per model, fetched 2026-07-17): gpt-oss-120b / zai-glm-4.7 / gemma-4-31b: 5 RPM, 30K TPM, 1M TPH, 1M TPD. cheahjs README (older): gpt-oss-120b: 30 requests/minute, 60,000 tokens/minute, 900 requests/hour, 1,000,000 tokens/hour, 14,400 requests/day, 1,000,000 tokens/day." Category: permanent_free — but the gotcha: the official rate-limits page now labels the tier "Free Trial" and shows 5 RPM, far below the 30 RPM the seed repos reported in June 2026. Treat the official 5 RPM as current, and watch for the tier turning into an expiring trial.

Friction check
Credit card no
Phone verification unknown
Org/membership no
Est. time to first key 5m
Trains on your prompts unknown

Get a key

  1. Go to cloud.cerebras.ai and sign up (or log in).
  2. In the left navigation bar, click API Keys.
  3. Create a key and copy it (console path /console/api-keys).
  4. export CEREBRAS_API_KEY="your-api-key-here" (the docs' own convention).

First request

Model and curl confirmed against the official quickstart. Success shape: JSON with choices[0].message.content.

export CEREBRAS_API_KEY="YOUR_KEY"
curl https://api.cerebras.ai/v1/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer $CEREBRAS_API_KEY" \
  -d '{"model":"gpt-oss-120b","messages":[{"role":"user","content":"Say OK."}],"max_tokens":16}'

Limits that will bite you

  • Per model (catalog, official rate-limits page): 5 RPM, 30K TPM, 1M TPH, 1M TPD for gpt-oss-120b, zai-glm-4.7, and gemma-4-31b. Not a shared pool.
  • 5 RPM is very low — one burst of parallel calls will 429 (429 Too Many Requests per quickstart). Serialize requests.
  • gemma-4-31b (vision): 2 images per request, 4 MB payload (catalog).
  • Seed repos report an 8K context cap on the free tier — not confirmed on the official page (unconfirmed — verify with a long prompt).
  • Max output ~8,000 tokens on gpt-oss-120b / zai-glm-4.7 (catalog).
  • Conflict on record: seed repos said 30 RPM / 14,400 RPD; official page says 5 RPM. ModelHubby treats the official page as current.

Key rotation runbook

  • Second key: same console — API Keys → create another. Whether multiple simultaneous keys are allowed is unconfirmed (unconfirmed — verify on first signup).
  • Revoke: delete the key from the same API Keys page (exact button label unconfirmed).
  • ModelHubby integration: export CEREBRAS_API_KEY_2="..." — the key pool picks it up automatically; state machine cools down rate-limited keys per-key.

ModelHubby usage

pnpm mh show cerebras
pnpm mh probe --live --provider cerebras --apply