---
license: apache-2.0
base_model: Qwen/Qwen3.5-35B-A3B
tags:
  - trading
  - solana
  - defi
  - prediction-markets
  - grpo
  - moe
  - autonomous-agents
  - infraxa
language:
  - en
pipeline_tag: text-generation
library_name: transformers
model-index:
  - name: Qwen3.5-Trading-Agent
    results: []
---

# Qwen3.5-Trading-Agent

**By [Infraxa](https://infraxa.ai) — The Execution Layer for Autonomous Finance**

A GRPO-finetuned Qwen3.5-35B-A3B Mixture-of-Experts model, trained on Solana on-chain trading data and prediction market signals. Built for autonomous trade execution, swap routing, and market reasoning.

## Model Details

| Parameter | Value |
|---|---|
| Base Model | Qwen3.5-35B-A3B (MoE) |
| Architecture | `Qwen3_5MoeForCausalLM` |
| Total Parameters | ~35B |
| Active Parameters | ~3B per token |
| Experts | 256 total, 8 active per token |
| Hidden Size | 2048 |
| Layers | 40 (30 linear attention + 10 full attention) |
| Context Length | 262,144 tokens |
| Precision | bfloat16 |
| Training Method | GRPO (Group Relative Policy Optimization) |

## Training Data

This model was GRPO-trained on:

- **Solana on-chain transaction data** — real swap and trade executions across DEXs
- **Prediction market data** — outcomes, odds, and resolution signals
- **Trading run logs** — full execution traces including routing, slippage, and settlement

The training objective optimizes for accurate trade reasoning: identifying optimal swap routes, predicting market movements, and generating executable trade instructions.

## Intended Use

- Autonomous trading agents on Solana
- Swap execution and routing decisions
- Prediction market analysis and position sizing
- On-chain data interpretation and trade signal generation
- Integration with Infraxa's execution layer for gasless, agent-driven finance

## Architecture

Qwen3.5-35B-A3B uses a hybrid attention design with both linear and full attention layers in a 3:1 ratio. The MoE architecture (256 experts, 8 active) gives the model high capacity while keeping inference costs low — only ~3B parameters are active per forward pass.

## How to Use

```python
from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "infraxaai/Qwen3.5-Trading-Agent"

tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype="bfloat16",
    device_map="auto",
)

prompt = "Analyze the current SOL/USDC liquidity across Orca, Raydium, and Jupiter. Recommend the optimal swap route for 10,000 USDC."
messages = [{"role": "user", "content": prompt}]

text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=512)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))
```

## Links

- [Infraxa Platform](https://infraxa.ai)
- [Infraxa Docs](https://docs.infraxa.ai)
- [Infraxa Dashboard](https://dashboard.infraxa.ai)
- [Twitter](https://twitter.com/infraxa)

## License

Apache 2.0 — same as the base Qwen3.5 model.