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.
- Downloads last month
- -
Hardware compatibility
Log In to add your hardware
We're not able to determine the quantization variants.
Model tree for crymadxAI/CrymadX-LLM-32B
Evaluation results
- Tool Selection on crymadxAI/CryptoExec-Bench View evaluation results 92.4 *
- Anti Chatbot on crymadxAI/CryptoExec-Bench View evaluation results 96 *
- Social Engineering Refusal on crymadxAI/CryptoExec-Bench View evaluation results 97.5 *
- Multilingual Execution on crymadxAI/CryptoExec-Bench View evaluation results 87.5 *
- Voice Processing on crymadxAI/CryptoExec-Bench View evaluation results 90 *
- Image Ocr Processing on crymadxAI/CryptoExec-Bench View evaluation results 90 *
- Overall on crymadxAI/CryptoExec-Bench View evaluation results 91.8 *