Kairo v1.5 — Foundation Crypto-Native Language Model (32k Context)
Kairo v1.5 (kairo-crypto-model/kairo-v1.5) is an open-weights, crypto-native causal language model architecture pre-trained continuously on live web streams of technical documentation, protocol whitepapers, smart contracts (Anchor & Solidity), audit post-mortems, and governance proposals.
Funded directly by pump.fun creator fees of the $JELLY token, all compute is settled on Solana and verified with on-chain transaction receipts.
🌟 What's New in Kairo v1.5
- 32,768 Token Context Window: Expanded from 4,096 in v1.1. Ingest entire multi-contract DeFi protocols, complete Anchor IDLs, and full audit reports in a single inference call.
- 25,000+ Cleaned Crypto Documents: Pretrained on verified pages cleaned and deduplicated (SHA-256 + 64-bit SimHash Hamming distance) from the public
kairo-crypto-model/kairo-dataset. - Crypto Benchmark v1.5: Scored 86.4% across DeFi math, protocol specifications, tokenomics, and consensus reasoning (vs 62.0% for Kairo v1.1 and 24.5% for legacy 42M toy crawlers).
- On-Chain Compute Verification: Pre-training runs executed on 8x NVIDIA H100 SXM5 80GB GPU clusters with Solana transaction proofs posted to the public ledger.
📊 Benchmark Evaluation
| Dimension | Legacy 42M Models | Kairo v1.1 | Kairo v1.5 |
|---|---|---|---|
| Context Length | 1,024 tokens | 4,096 tokens | 32,768 tokens |
| Architecture | Toy nanochat (16.9M active) | 0.1B Native Causal LM | 0.1B Foundation LM + GQA |
| DeFi Math & AMMs | 18.2% | 68.5% | 88.2% |
| Anchor IDLs & SVM | 12.0% | 59.4% | 85.0% |
| Governance & EIPs | 31.4% | 61.2% | 84.8% |
| Overall Crypto Eval | 24.5% | 62.0% | 86.4% |
🚀 Quickstart
from transformers import AutoModelForCausalLM, AutoTokenizer
model_id = "kairo-crypto-model/kairo-v1.5"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype="bfloat16")
prompt = "Explain Solana stake account delegation and warm-up/cool-down cooldown mechanics in Anchor:"
inputs = tokenizer(prompt, return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=512, temperature=0.2)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))
🔗 Verifiable Provenance
- Dataset: kairo-crypto-model/kairo-dataset
- Live Swarm Screencasts: kairo.trade
- Solana Treasury:
3VfQ15ijXXsevehQDmJPrK3sRXcqHQifKLGYFqYd3VqW - $JELLY Mint:
CyDJMENR6Sbtwj4erhuARNa5cNbeK92GpNAGyXLVpump - License: Apache 2.0
- Downloads last month
- -