Exeaon1-Claw-32B 🦅🔥

Exeaon1-Claw-32B is the flagship coder model powering the ExeaonClaw autonomous coding agent and Exeaon Canvas.

Compressed from Qwen/Qwen2.5-Coder-32B-Instruct using ℰ-PURE (Sine ℰ-Grid Lloyd-Max quantization with column-wise error compensation and static rANS entropy coding).

Runs natively with epure-runtime, and stays compressed resident in memory — zero reload latency between agentic turns.

Metric	Specification
Base Model	Qwen/Qwen2.5-Coder-32B-Instruct
Original Size	61.2 GB (fp16)
Compressed Size	16.363 GB
Compression Ratio	3.73×
Bits Per Weight	4.30 bpw (measured index entropy)
Calibration	128 sequences, 64/64 Transformer Blocks error-compensated
Format	.ebin (entropy-coded container)

---

⚡ Multi-Agent Ecosystem Architecture

Exeaon1-Claw-32B is designed to serve as the unified intelligence substrate across the Exeaon stack:

Exeaon Canvas (Native Desktop) ➔ LangGraph ➔ {CrewAI, AI-Q, Goose} ➔ NeMo Tracing
                                       ↓
                        The One OpenAI Seam (agent_serve.py)
                                       ↓
                           Exeaon1-Claw-32B (.ebin)

• Strict Tool Calling: Native OpenAI function calling schema support (bash execution, file editing, diff manipulation).
• Resident Mmap Memory Cache: Load once (~2.6s), execute 200+ multi-turn steps with 0ms reload overhead.

---

🚀 Quickstart

1. Run with epure-runtime:

pip install epure-runtime
python -m epure.runtime Exeaon/Exeaon1-Claw-32B "Write a quicksort function in Rust:"

2. Serve to Exeaon Canvas / Agent Frameworks:

python src/agent_serve.py Exeaon/Exeaon1-Claw-32B --port 8000