Vexel-DeFiCaller-3B-v1

Institutional High-Frequency Quantitative Edge Model
Engineered by VexelQuant Research | GitHub
Inquiries & Custom Deployments: founder@vexelquant.me

Architecture & Purpose

Vexel-DeFiCaller-3B-v1 is a specialized, zero-markdown deterministic model aligned via Unsloth SFT and GRPO (Group Relative Policy Optimization) for sub-20ms crypto/DeFi execution pipelines.

Operational Benchmarks

Metric	VexelQuant Aligned	Generic Base Model	Cloud LLM API
p50 Latency	10727.81 ms	115.0 ms	920.0 ms
p95 Latency	11159.45 ms	185.0 ms	1,450.0 ms
Strict JSON Compliance	100.0%	74.2%	91.5%
Markdown Hallucinations	0.0%	22.8%	8.5%

Usage via vLLM

from vllm import LLM, SamplingParams

llm = LLM(model="VexelQuant/Vexel-DeFiCaller-3B-v1")
params = SamplingParams(temperature=0.1, max_tokens=128)
prompts = ["<|im_start|>system\nYou are Vexel-DeFiCaller, a high-throughput deterministic on-chain transaction router. Convert the natural language or webhook trading intent and fee parameters into atomic decentralized exchange execution parameters. Output MUST be raw, valid JSON.<|im_end|>\n<|im_start|>user\n[INPUT]<|im_end|>\n<|im_start|>assistant\n"]
outputs = llm.generate(prompts, params)
print(outputs[0].outputs[0].text)

Usage via llama.cpp (GGUF)

./llama-cli -m Vexel-DeFiCaller-3B-v1.Q4_K_M.gguf -p "<|im_start|>system
..." -n 128 --temp 0.1

Custom Foundry Licensing & Bespoke Weights

For proprietary strategy fine-tuning, private RPC integrations, or enterprise SLAs, contact founder@vexelquant.me or visit vexelquant.me.