Find everything you need to build your project. Use these platform links, AI assistants, and machine learning toolkits to accelerate your building cycle.

⭐ Featherless.ai — Free Plan for GIBC Builders [Sponsor Perk]:

Featherless.ai, our official partner, is giving every GIBC participant one month free — access to 30,000+ open-source AI models (DeepSeek, Kimi, GLM, Mistral, Qwen and more) through a single API. No GPUs to manage, no per-token billing.

👉Join to get the code in the starter kit

  • How redemption works: The link opens a sign-up page first — create your account, and you will then be taken to the subscription page with the discount already applied. If the discount does not appear, reload once. Still missing? Open the redeem link again from scratch rather than navigating there yourself — the code only travels with that specific link.
  • Before you confirm: Set a calendar reminder to cancel before the free month ends if you do not want to continue. This is a real subscription, not a trial credit.
  • Get your API key: Profile icon (top right) → API Keys → create → copy. It starts with fw-. Treat it like a password — never paste it straight into your code and never commit it to GitHub. Use an environment variable or a .env file listed in your .gitignore.

Using it — it speaks the same language as OpenAI. Any OpenAI tutorial or library works; you only change one line:

import os
from openai import OpenAI

client = OpenAI(
    api_key=os.environ["FEATHERLESS_API_KEY"],
    base_url="https://api.featherless.ai/v1"   # <-- the only change
)

response = client.chat.completions.create(
    model="deepseek-ai/DeepSeek-V3.2",
    messages=[{"role": "user", "content": "Hello!"}]
)

print(response.choices[0].message.content)

The same trick works in JavaScript/Node, LangChain, the Vercel AI SDK and Cursor — anywhere you can set a base URL or custom endpoint, point it at https://api.featherless.ai/v1

Recommended models (copy the ID exactly as written):

  • Reasoning & coding: deepseek-ai/DeepSeek-V3.2
  • Agents & tool use: MiniMaxAI/MiniMax-M2.5
  • Very long documents: moonshotai/Kimi-K2.5
  • Long multi-step tasks: zai-org/GLM-5
  • Fast & lightweight: mistralai/Mistral-Nemo-Instruct-2407

⚡ Important — how the limits actually work: Featherless does not charge per token, so send as many tokens as you like. Instead your plan limits how many run at the same time. You get 4 units: models under 24B cost 1 unit each (4 at once), models from 24B to 34B cost 2 units (2 at once), and models above 34B cost all 4 units (one at a time).

This matters for your architecture: every big-name model above costs all 4 units, so if your project processes things in parallel it will queue up and feel slow. The fix: use a small model for the bulk work and save the big model for the hard part — mistralai/Mistral-Nemo-Instruct-2407 costs 1 unit, so you can run 4 at once.

Common errors: 401 — key not recognised, re-copy it or make a new one. 403 — model is locked, open its page and click "Unlock Model". 429 — out of concurrency units, wait and retry. 500 — you sent a setting the model does not support. 503 — model waking up or at capacity, retry; if it fails three times ask in the GIBC Discord.

Browse all models at featherless.ai/models · Full documentation at featherless.ai/docs

Essential Communications & Platforms:

  • Devpost [Required]: The official submission and hacking platform. Track progress and submit your final project here.

  • Discord [Recommended]: The central hub for GIBC V2. Join to find teammates, receive real-time mentor support, and track live announcements.

AI Assistants & Development Tools:

  • Cursor: An AI-first code editor with deep repository context—highly recommended for full-stack prototyping.

  • Claude Code: Advanced AI-assisted command-line coding tool to accelerate development pipelines.

  • OpenCode: An AI-powered CLI tool built specifically for rapid interface design and system prototyping.

  • ChatGPT / OpenAI: Access standard LLM APIs and leverage a robust playground environment for quick prototyping.

  • GitHub Copilot: Your AI pair programmer directly integrated within VS Code to speed up boilerplate code generation.

  • v0 by Vercel: Generate clean, modern frontend UI components in seconds using simple text prompts.

  • Replit: A fully collaborative web-based cloud IDE featuring a built-in AI agent to code and deploy instantly.

Machine Learning, Training & Hardware Hubs:

  • Hugging Face: The open-source standard. Access pre-trained foundational models, curated datasets, and open model training hubs.

  • Unsloth: A memory-efficient LLM fine-tuning library that allows you to train small-scale models up to 2x faster.

  • Weights & Biases: Seamless experiment tracking, hyperparameter logging, and loss curve visualization.

  • Groq: An ultra-fast, low-latency LLM inference engine ideal for powering live application backends.

  • Perplexity: An AI-powered research engine featuring inline citations to assist with literature and market review.

Frontend, Design & Database Hosting:

  • Figma: The premier collaborative interface layout tool for software/app UI/UX design and wireframing.

  • Vercel: Fast hosting for modern frontend applications and lightweight serverless backend functions.

  • Netlify: Seamless static site deployment, forms integration, and continuous delivery setups.

  • Supabase: An open-source, relational Firebase alternative providing quick Postgres databases, authentication, and instant APIs.

Frameworks & Learning Foundations:

  • Fast.ai: A widely respected learning platform providing practical, code-first deep learning courses from the ground up.

  • TensorFlow: Google's production-ready, open-source deep learning framework for building data-driven systems.

  • JAX: A high-performance, hardware-accelerated machine learning framework optimized for complex, custom transformer architectures.

Contact Us & Support Channels:

If you have any questions, need technical guidance, or require assistance with team formation, please connect with the organizing team directly through our official channels:

Note for Track 01: pretrained models and fine-tuning are not permitted in the LLM track. Hugging Face and Unsloth are useful for tokenizers, datasets, and evaluation, not for your model weights. See the Rules page.