crymadxAI/CrymadX-LLM-32B

CrymadX LLM 32B

An autonomous crypto execution agent fine-tuned from Qwen 2.5 32B-Instruct. Built to EXECUTE crypto operations through tool calls — not give instructions.

What Makes This Different

Most AI assistants tell you how to do things. CrymadX LLM does them for you:

User: check my BTC balance
AI: [calls get_balance(token="BTC")] → You have 0.5432 BTC (~$35,310)

User: swap 0.1 ETH to USDT  
AI: [calls get_swap_estimate()] → 0.1 ETH → 329.67 USDT. Confirm?
User: yes
AI: [calls execute_swap()] → Done! 0.1 ETH → 329.67 USDT

Model Details

Field	Value
Base Model	Qwen 2.5 32B-Instruct
Fine-tuning	QLoRA (rank 64) via Unsloth
Training Data	11,606 examples
Training Time	17.1 hours on NVIDIA A40
Final Loss	0.0558
License	Apache 2.0

Available Quantizations

File	Size	VRAM	Use Case
	19 GB	24 GB	Single RTX 3090 / RTX 4090
	33 GB	48 GB	2x RTX 3090 / A40 / A6000

37 Tools Available

The model is trained to call tools across these categories:

• Wallet: get_balance, get_all_balances, get_deposit_address, get_transactions
• Trading: get_price, get_swap_estimate, execute_swap, create_price_alert, create_auto_invest
• Fiat: create_fiat_buy_order, create_fiat_sell_order
• Transfers: validate_address, estimate_send_fee, preview_transaction, request_auth, execute_send
• Staking: stake_asset, unstake_asset, get_staking_positions
• Savings: create_vault, unlock_vault, get_vault_positions
• Account: update_profile, start_kyc, setup_2fa, verify_and_enable_2fa
• Card: fund_card, get_card_balance
• Support: create_support_ticket
• Market: get_market_overview

Training Data Composition

Category	Examples	Description
Single-tool operations	~5,000	Balance, price, swap, send, buy, sell, stake
Multi-turn chains (5-20 turns)	~3,500	Full sessions with noise, interruptions, errors
Beginner/naive users	~600	Elderly, teens, typos, wrong terminology
Multilingual	~600	Dutch, French, Spanish, German, Arabic, Turkish, Portuguese, Pidgin
Error recovery	~500	Insufficient balance, wrong address, wrong OTP
Social engineering refusal	~400	Prompt injection, jailbreak attempts
Media handling	~400	Voice transcripts, OCR/QR images, stickers
Anti-chatbot guardrails	~300	Execute, never instruct

Languages

Trained and tested in: English, Dutch, French, Spanish, German, Arabic, Turkish, Portuguese, Nigerian Pidgin, Indonesian, Swahili, and mixed-language conversations.

Anti-Chatbot Design

The model is explicitly trained to NEVER give instructions:

• "How do I check my balance?" → Calls immediately
• "How do I swap?" → Asks which tokens, then executes
• "How do I buy BTC?" → Asks how much, then creates buy order

Usage with vLLM

python -m vllm.entrypoints.openai.api_server \
    --model crymadxAI/CrymadX-LLM-32B \
    --quantization gguf \
    --max-model-len 8192

Built By

CrymadX — Crypto Exchange Platform

Fine-tuned with Unsloth for 2x faster training.