Free LLM API Resources (reference + harvest)

Source: cheahjs/free-llm-api-resources (curated list of free LLM inference APIs). Registered 2026-07-18. This is a reference/watchlist, not a capability by itself. Value = cost-saving map: free tiers for background/batch/non-critical tasks.

🔴 Red line

Free tiers are rate-limited + not flagship accuracy → NEVER enter the golden fallback chain (accuracy first). Use for batch / background / experiments only.

Already in our arsenal (keys present)

Google/Gemini (dual key) · Groq · Mistral · OpenRouter · HuggingFace.

✅ Already reachable via OpenRouter (NO new key needed — verified live 2026-07-18)

We have OPENROUTER_API_KEY, and OpenRouter fronts BOTH targets:

• Cohere → cohere/command-a (multilingual incl Arabic — tested live ✅) · cohere/command-r7b-12-2024 · cohere/command-r-plus-08-2024.
• Cerebras (ultra-fast) → route via "provider":{"order":["Cerebras"]} — tested live ✅ on openai/gpt-oss-120b.

So Cerebras speed + Cohere are usable RIGHT NOW through OpenRouter; no signup. Arabic-specialized SKU command-r7b-arabic-02-2025 is Cohere-direct only (not on OpenRouter). ✅ COHERE_API_KEY IS NOW SET in env.sh (2026-07-18) — Arabic SKU tested live, works. No further action needed.

Direct-key providers (optional — only if you want native/free-tier direct)

| Provider | Why | Free limits | Signup | |---|---|---|---| | Cerebras | world's fastest inference (background speed) | free tier, key from console | cloud.cerebras.ai | | Cohere | command-r7b-arabic (Arabic-tuned) + Command A | trial key: 1,000 calls/mo, 20 req/min | dashboard.cohere.com/api-keys | | SambaNova | fast free open models | per-model limits | cloud.sambanova.ai | | NVIDIA NIM | free open models | 40 req/min | build.nvidia.com | | GitHub Models | free open models | account-based | github.com/marketplace/models |

Activation (once you paste a key)

Add to /data/.openclaw/secrets/env.sh:

bash
export CEREBRAS_API_KEY="..."   # from cloud.cerebras.ai
export COHERE_API_KEY="..."     # from dashboard.cohere.com/api-keys

Then test:

bash
. /data/.openclaw/secrets/env.sh
/usr/bin/python3 skills/free-llm-resources/cerebras_client.py "مرحبا"
/usr/bin/python3 skills/free-llm-resources/cohere_client.py "ترجم: hello world"

Both are OpenAI-compatible-ish; scripts self-check for the key and print how to get it.

✅ Unified router + Portal integration (2026-07-18)

• Internal capability: free_llm.py "prompt" [--backend cohere-arabic|command-a|cerebras|auto] [--doc PATH] [--json]

- never-stop fallback across the 3 backends; extracts doc text (txt/md/pdf/docx/csv) as context (≤ 24k chars). - cohere-arabic = Cohere direct (Arabic SKU) · command-a = OpenRouter cohere/command-a · cerebras = OpenRouter@Cerebras (gpt-oss-120b).

• Portal (external): academic portal /free-llm page (dashboard card “🤝 الموارد المجانية”).

- Form: backend selector + prompt + optional document upload (multipart, ≤ 8MB) → answered by free_llm.py. - Tested live E2E (2026-07-18): text + document upload both work across all 3 backends.

Placement in arsenal

• Registered in scripts/weekly_arsenal_audit.sh watchlist (monthly re-check for new free sources).
• When keys are active: Cerebras → fast background/batch text; Cohere → Arabic-tuned cheap tasks.