← knowledge.oriz.in

@chirag127/oriz-ai-providers (18th package) + chirag127/oriz-ai-providers-data data repo

decision decisionpackageaiprovidersfree-tierdata-repo

@chirag127/oriz-ai-providers (18th package)

Decision

Add @chirag127/oriz-ai-providers as the 18th package in the family. Its job: a thin wrapper around every free LLM API the family uses for blog rewrites, omni-publish drafts, janaushdhi substitute-finder, ncert summaries, etc.

Why a package + data-repo split:

Providers list (as of 2026-06-22)

Tier 1: Anonymous (no key, no signup, no card)

  1. OVHcloud AI Endpoints — 2 RPM per IP per model, EU-hosted, 20+ models (Qwen3.5, gpt-oss, Llama 3.3, Mistral)
  2. LLM7.io — 30 RPM per IP, 30+ models (deepseek-r1, gpt-4o-mini, gemini-2.5-flash-lite, etc.)
  3. Pollinations — anonymous gpt-oss-20b

Tier 2: Free with no-card signup

  1. Cerebras — 30 RPM + 1M TPD, ultra-fast (~2,600 tok/s), gpt-oss-120b + Llama 3.1 8B
  2. Groq Cloud — 30 RPM + 1,000 RPD, llama-3.3-70b-versatile (faster than NIM)
  3. NVIDIA NIM — 40 RPM, 100+ models, requires phone verification
  4. Google AI Studio — Gemini 2.5 / 3.x Flash, 5-15 RPM + 20-1,500 RPD per model (free outside EU/UK/Switzerland)
  5. Cohere — 20 RPM + 1,000 req/month, Command A+ / R+ (non-commercial only)
  6. GitHub Models — 10-15 RPM + 50-150 RPD, GPT-5 / GPT-4.1 / o4-mini, free with Copilot tier
  7. Cloudflare Workers AI — 10K neurons/day, Llama 3.3 70B FP8 / GPT-OSS / Qwen3
  8. HuggingFace — 100K credits/mo, router to Fireworks/Together/Hyperbolic, thousands of models
  9. Mistral La Plateforme — 500K TPM + ~1B tokens/month (Experiment plan), Mistral Medium 3.5 / Codestral
  10. SambaNova — 20 RPM + 200K TPD, DeepSeek V3.1 + Llama 3.3 70B
  11. OpenRouter — 20 RPM + 200 RPD per :free model (Llama 3.3 70B, Qwen3-Coder, Nemotron-Ultra-550B)
  12. Z.AI (Zhipu) — GLM-4.7-Flash + GLM-4.6V-Flash (Chinese provider)
  13. SiliconFlow — Qwen3-8B + DeepSeek-R1-Distill (Chinese)
  14. Aion Labs — 15 RPM + 20K TPD, roleplay-specialized
  15. Ollama Cloud — qualitative usage, 400+ Ollama-hosted models (not OpenAI SDK)
  16. ModelScope — 2,000 RPD, Qwen3.5-35B-A3B + Qwen3.5-27B (requires Alibaba real-name)
  17. Kilo Code — auto-router free models

Priority order (default fallback chain)

For text completion at low rate:

  1. OVHcloud anonymous (zero friction, EU-hosted)
  2. LLM7 anonymous
  3. Cerebras (key required, ultra-fast)
  4. Groq Cloud (key, also fast)
  5. NVIDIA NIM (key + phone verified, more model variety)
  6. OpenRouter free (key, broad coverage)
  7. Google AI Studio Gemini (key, Gemini-flavored output)
  8. CF Workers AI (key, lives in our infra)

For high-volume (>30 RPM):

For reasoning tasks:

For vision/multimodal:

Data repo shape

chirag127/oriz-ai-providers-data:

providers.json    # one entry per provider
models.json       # one entry per model (with provider link)
rate-limits.json  # provider × model × tier
env-vars.json     # which env var maps to which provider
signup-urls.json  # for the README + onboarding doc
priority.json     # default fallback chain (the order above)

Updated via PR. Each change creates a new release tag. Package fetches latest tag (or main) at build/runtime.

Wrapper API

import { ai } from "@chirag127/oriz-ai-providers";

const result = await ai.complete({
  prompt: "Rewrite this blog post for Twitter",
  task: "rewrite-short",  // mapped to priority chain
  maxTokens: 280,
  // optional overrides:
  preferProvider: "cerebras",
  fallback: true,
});

// result.text, result.provider, result.model, result.tokensUsed

The wrapper:

  1. Loads provider data from data-repo (cached 24h)
  2. Picks the highest-priority provider with a configured env var
  3. Calls it via OpenAI SDK (most providers are OpenAI-compatible)
  4. On 429/5xx: falls back to next provider
  5. Returns first successful result

Master pointer

Adding this package brings the family count to 18 packages (was 17 per the-23-packages.md). Rename + update count in:

Supersedes in part

decisions/architecture/stack-picks-2026-06-22.md — its "AI inference" section named NIM primary + OpenRouter fallback only. That's now superseded by this decision (priority chain in this file). Update that file to point here.

Cross-refs