Less is More: Recursive Reasoning with Tiny Networks
Paper
•
2510.04871
•
Published
•
507
🧠 SOLANA $OS
The Neural Operating System for Solana
██████╗ ███████╗
██╔═══██╗██╔════╝
██║ ██║███████╗
██║ ██║╚════██║
╚██████╔╝███████║
╚═════╝ ╚══════╝
Recursive Intelligence • Token-Gated Inference • Autonomous Agents
💎 Token Contract Address
DrU9M6SUaXWua49zeaHQWJuwMpcZ4jMDRT3J5Ywpump
🌐 What is Solana $OS?
Solana $OS is a token-gated AI infrastructure layer combining a 3.5M parameter Tiny Recursive Model (TRM) with blockchain-native utility. It powers intelligent Solana development assistance, autonomous agents, and on-chain AI inference.
┌─────────────────────────────────────────────────────────────────┐
│ SOLANA $OS ECOSYSTEM │
├─────────────────────────────────────────────────────────────────┤
│ │
│ ┌─────────────┐ ┌─────────────┐ ┌─────────────┐ │
│ │ $OS TRM │───▶│ CLAWDBOT │───▶│ INFERENCE │ │
│ │ Model │ │ Agent │ │ API │ │
│ └─────────────┘ └─────────────┘ └─────────────┘ │
│ │ │ │ │
│ └──────────────────┼──────────────────┘ │
│ ▼ │
│ ┌─────────────┐ │
│ │ $OS TOKEN │ │
│ │ UTILITY │ │
│ └─────────────┘ │
│ │
└─────────────────────────────────────────────────────────────────┘
💰 $OS Token Utility
The $OS token unlocks access to the entire Solana OS ecosystem:
🎫 Tier System
| Tier | $OS Holdings | Benefits |
|---|---|---|
| 🥉 Bronze | 10,000 $OS | 100 inference calls/day |
| 🥈 Silver | 100,000 $OS | 1,000 inference calls/day + Clawdbot basic |
| 🥇 Gold | 500,000 $OS | 5,000 inference calls/day + Clawdbot pro |
| 💎 Diamond | 1,000,000+ $OS | Unlimited + Clawdbot autonomous + API keys |
🔑 Access Methods
1. Hold to Access (Recommended) Simply hold $OS tokens in your connected wallet to unlock tier benefits automatically.
2. Burn for Credits Burn $OS tokens to receive permanent inference credits:
- 1,000 $OS burned = 500 inference credits (never expire)
3. Stake for Premium Stake $OS to earn yield while maintaining access:
- Staked tokens count toward tier qualification
- Earn additional $OS from protocol fees
🤖 Clawdbot Integration
Clawdbot is an autonomous AI agent powered by Solana $OS, capable of:
// Clawdbot Capabilities
interface ClawdbotFeatures {
// Solana Development
codeAnalysis: "Smart contract review & optimization";
pdaGeneration: "Generate PDAs with derivation paths";
txBuilder: "Construct & simulate transactions";
// Autonomous Actions
portfolioMonitor: "Track & alert on positions";
dexAggregation: "Find optimal swap routes";
yieldFarming: "Auto-compound strategies";
// Intelligence
marketAnalysis: "Real-time sentiment & signals";
contractAudit: "Security vulnerability detection";
documentation: "Auto-generate project docs";
}
Clawdbot Access Tiers
| Feature | Silver | Gold | Diamond |
|---|---|---|---|
| Q&A Responses | ✅ | ✅ | ✅ |
| Code Analysis | ✅ | ✅ | ✅ |
| Transaction Building | ❌ | ✅ | ✅ |
| Autonomous Monitoring | ❌ | ❌ | ✅ |
| Custom Agent Training | ❌ | ❌ | ✅ |
| API Access | ❌ | ❌ | ✅ |
🧠 Model Architecture
Solana $OS is built on the Tiny Recursive Model (TRM) architecture:
┌────────────────────────────────────────────────────────┐
│ TRM-SOLANA-OS ARCHITECTURE │
├────────────────────────────────────────────────────────┤
│ │
│ Input: "What is a PDA?" │
│ │ │
│ ▼ │
│ ┌─────────────────────────────────────────────────┐ │
│ │ BYTE-LEVEL TOKENIZATION │ │
│ │ (UTF-8 → 258 vocab tokens) │ │
│ └─────────────────────────────────────────────────┘ │
│ │ │
│ ▼ │
│ ┌─────────────────────────────────────────────────┐ │
│ │ RECURSIVE REASONING CORE │ │
│ │ ┌─────────────────────────────────────────┐ │ │
│ │ │ H-Cycle 1 ──▶ H-Cycle 2 │ │ │
│ │ │ │ │ │ │ │
│ │ │ ▼ ▼ │ │ │
│ │ │ L-Cycle 1,2 L-Cycle 1,2 │ │ │
│ │ └─────────────────────────────────────────┘ │ │
│ └─────────────────────────────────────────────────┘ │
│ │ │
│ ▼ │
│ Output: Refined Solana expertise │
│ │
└────────────────────────────────────────────────────────┘
Specifications
| Component | Value |
|---|---|
| Parameters | 3.5M (~1/1000th of GPT-3) |
| Layers (L) | 1 transformer layer |
| High-level cycles (H) | 2 reasoning iterations |
| Low-level cycles (L) | 2 refinement iterations |
| Vocabulary | 258 tokens (byte-level) |
| Max Sequence | 512 tokens |
| Inference | CPU/MPS/GPU compatible |
| Memory | <1GB during inference |
🚀 Quick Start
1. Connect Wallet & Verify Holdings
import { Connection, PublicKey } from '@solana/web3.js';
import { getAssociatedTokenAddress, getAccount } from '@solana/spl-token';
const OS_TOKEN_MINT = new PublicKey('DrU9M6SUaXWua49zeaHQWJuwMpcZ4jMDRT3J5Ywpump');
async function checkOSBalance(walletPubkey: PublicKey): Promise<number> {
const connection = new Connection('https://api.mainnet-beta.solana.com');
const ata = await getAssociatedTokenAddress(OS_TOKEN_MINT, walletPubkey);
try {
const account = await getAccount(connection, ata);
return Number(account.amount) / 1e6; // Assuming 6 decimals
} catch {
return 0;
}
}
function getTier(balance: number): string {
if (balance >= 1_000_000) return 'diamond';
if (balance >= 500_000) return 'gold';
if (balance >= 100_000) return 'silver';
if (balance >= 10_000) return 'bronze';
return 'none';
}
2. API Inference
// Token-gated inference endpoint
const response = await fetch('https://api.solana-os.ai/v1/inference', {
method: 'POST',
headers: {
'Content-Type': 'application/json',
'X-Wallet-Address': walletAddress,
'X-Signature': signedMessage, // Prove wallet ownership
},
body: JSON.stringify({
question: 'How do I create a PDA in Anchor?',
model: 'trm-solana-os-v1',
max_tokens: 512,
}),
});
const { answer, tokens_used, remaining_calls } = await response.json();
3. Clawdbot Discord/Telegram
# Discord Commands (Silver+ tier)
/clawdbot ask "What is rent exemption in Solana?"
/clawdbot analyze <contract_address>
/clawdbot build-tx transfer 1 SOL to <address>
# Telegram Commands
/ask What are CPIs?
/audit <program_id>
/monitor <token_address>
🔧 Self-Hosted Inference
For Diamond tier holders, run the model locally:
Installation
# Clone the repository
git clone https://github.com/8bitlabs/solana-os-model
cd solana-os-model
# Install dependencies
pip install torch transformers huggingface_hub numpy
# Download model
python scripts/download_model.py --token YOUR_HF_TOKEN
Local Inference
import torch
import numpy as np
from huggingface_hub import hf_hub_download
from models.trm import TinyRecursiveReasoningModel
# Download checkpoint
checkpoint_path = hf_hub_download(
repo_id="ordlibrary/trm-solana-v1",
filename="model.pt"
)
# Load model
checkpoint = torch.load(checkpoint_path, map_location='cpu')
model = TinyRecursiveReasoningModel(**checkpoint['config']['arch'])
model.load_state_dict(checkpoint['model_state_dict'])
model.eval()
# Byte-level encoding
def encode(text: str, max_len: int = 512) -> torch.Tensor:
bytes_arr = np.frombuffer(text.encode('utf-8'), dtype=np.uint8)
tokens = (bytes_arr + 2).astype(np.uint8)
seq = np.zeros(max_len, dtype=np.uint8)
seq[:min(len(tokens), max_len-1)] = tokens[:max_len-1]
seq[min(len(tokens), max_len-1)] = 1 # EOS
return torch.tensor(seq).unsqueeze(0)
def decode(tokens: torch.Tensor) -> str:
tokens = tokens.squeeze().numpy()
tokens = tokens[tokens > 1] - 2 # Remove PAD/EOS, unshift
return bytes(tokens).decode('utf-8', errors='replace')
# Inference
question = "What is a Program Derived Address (PDA) in Solana?"
input_tensor = encode(question)
with torch.no_grad():
output = model(input_tensor)
answer = decode(output.argmax(dim=-1))
print(answer)
📊 Training Details
Dataset
| Metric | Value |
|---|---|
| Total Pairs | 8,970 Q&A |
| Training Set | 8,073 (90%) |
| Test Set | 897 (10%) |
| Topics | Solana architecture, accounts, programs, transactions, security, DeFi |
| Encoding | Byte-level UTF-8 |
Training Configuration
# config/training.yaml
architecture:
L_layers: 1
H_cycles: 2
L_cycles: 2
vocab_size: 258
max_seq_len: 512
training:
epochs: 5000
global_batch_size: 64
learning_rate: 1e-4
weight_decay: 0.5
ema_enabled: true
ema_rate: 0.999
hardware:
device: mps # Apple Silicon
training_time: "~4 hours"
Performance
| Metric | Value |
|---|---|
| Training Loss | ~2.3 |
| Test Loss | ~2.5 |
| Inference Speed | <100ms |
| Memory Usage | <1GB |
🎯 Example Queries
The Solana $OS model excels at:
┌────────────────────────────────────────────────────────┐
│ EXAMPLE QUERIES │
├────────────────────────────────────────────────────────┤
│ │
│ ✓ "What is a Program Derived Address (PDA)?" │
│ ✓ "How do Solana transactions differ from Ethereum?" │
│ ✓ "Explain the Solana account model" │
│ ✓ "What is rent exemption?" │
│ ✓ "How do cross-program invocations (CPI) work?" │
│ ✓ "What is the System Program?" │
│ ✓ "How does Anchor simplify Solana development?" │
│ ✓ "What are the security best practices for PDAs?" │
│ │
└────────────────────────────────────────────────────────┘
🛡️ Tokenomics
┌─────────────────────────────────────────────────────────┐
│ $OS TOKENOMICS │
├─────────────────────────────────────────────────────────┤
│ │
│ Token: $OS │
│ Contract: DrU9M6SUaXWua49zeaHQWJuwMpcZ4jMDRT3J5Ywpump │
│ Chain: Solana │
│ Standard: SPL Token │
│ │
│ ┌─────────────────────────────────────────────────┐ │
│ │ UTILITY FLYWHEEL │ │
│ │ │ │
│ │ ┌──────────┐ │ │
│ │ │ HOLD │◀─────────────────┐ │ │
│ │ │ $OS │ │ │ │
│ │ └────┬─────┘ │ │ │
│ │ │ │ │ │
│ │ ▼ │ │ │
│ │ ┌──────────┐ ┌─────┴────┐ │ │
│ │ │ ACCESS │ │ VALUE │ │ │
│ │ │ TIERS │ │ ACCRUAL │ │ │
│ │ └────┬─────┘ └──────────┘ │ │
│ │ │ ▲ │ │
│ │ ▼ │ │ │
│ │ ┌──────────┐ ┌─────┴────┐ │ │
│ │ │ USE │───────────▶│ DEMAND │ │ │
│ │ │ SERVICES│ │ GROWS │ │ │
│ │ └──────────┘ └──────────┘ │ │
│ │ │ │
│ └─────────────────────────────────────────────────┘ │
│ │
│ Revenue Streams: │
│ • Inference API fees → 50% to stakers │
│ • Clawdbot premium features → 30% buyback & burn │
│ • Enterprise licensing → Treasury │
│ │
└─────────────────────────────────────────────────────────┘
🔗 Integration Examples
Next.js App Router
// app/api/inference/route.ts
import { NextRequest, NextResponse } from 'next/server';
import { verifyWalletSignature, checkTier } from '@/lib/auth';
export async function POST(req: NextRequest) {
const { question, walletAddress, signature } = await req.json();
// Verify wallet ownership
const isValid = await verifyWalletSignature(walletAddress, signature);
if (!isValid) {
return NextResponse.json({ error: 'Invalid signature' }, { status: 401 });
}
// Check $OS holdings
const tier = await checkTier(walletAddress);
if (tier === 'none') {
return NextResponse.json({
error: 'Insufficient $OS holdings',
required: 10000,
contract: 'DrU9M6SUaXWua49zeaHQWJuwMpcZ4jMDRT3J5Ywpump'
}, { status: 403 });
}
// Process inference
const answer = await runInference(question);
return NextResponse.json({ answer, tier });
}
React Hook
// hooks/useSolanaOS.ts
import { useWallet } from '@solana/wallet-adapter-react';
import { useState } from 'react';
export function useSolanaOS() {
const { publicKey, signMessage } = useWallet();
const [loading, setLoading] = useState(false);
const askQuestion = async (question: string) => {
if (!publicKey || !signMessage) {
throw new Error('Wallet not connected');
}
setLoading(true);
try {
// Sign message to prove ownership
const message = new TextEncoder().encode(`Solana OS: ${Date.now()}`);
const signature = await signMessage(message);
const response = await fetch('/api/inference', {
method: 'POST',
headers: { 'Content-Type': 'application/json' },
body: JSON.stringify({
question,
walletAddress: publicKey.toBase58(),
signature: Buffer.from(signature).toString('base64'),
}),
});
return await response.json();
} finally {
setLoading(false);
}
};
return { askQuestion, loading };
}
⚠️ Limitations
| Limitation | Description |
|---|---|
| Domain Specific | Trained only on Solana content |
| Knowledge Cutoff | Limited to training data timeframe |
| Not for Audits | Do not use for production security audits |
| English Only | Trained exclusively on English data |
| Q&A Format | Optimized for questions, not conversations |
📜 License
MIT License - Open source and free to use.
🙏 Acknowledgments
- Alexia Jolicoeur-Martineau - TRM Architecture
- Solana Foundation - Documentation & Ecosystem
- OrdLibrary - Original model training
- 8bit Labs - $OS Token Integration
📚 Resources
| Resource | Link |
|---|---|
| 📄 TRM Paper | arXiv:2510.04871 |
| 💻 TRM Code | GitHub |
| 🤗 Model | HuggingFace |
| 🌐 Solana Docs | docs.solana.com |
| 💎 Buy $OS | Pump.fun |
💎 Get $OS Now
DrU9M6SUaXWua49zeaHQWJuwMpcZ4jMDRT3J5Ywpump
Pump.fun • DexScreener • Birdeye
Built by 8bit Labs for the Solana ecosystem
The Neural Operating System for Solana