aiqlick/1.3 โ€” Frontier Autonomous Multimodal Recruitment Agent

aiqlick/1.3 is a production-grade 4B-parameter autonomous agent specialized in candidate search, CV/JD parsing, multidimensional candidate evaluation, and automated recruitment workflows.

Key Highlights

  • Base Architecture: Qwen/Qwen3.5-4B with Gated DeltaNet linear attention (hybrid linear + full attention).
  • Two-Stage Fine-Tuning:
    1. Stage 1 (SFT): Supervised fine-tuning on high-quality recruitment trajectory datasets validated with Gemini 3.7 Flash.
    2. Stage 2 (DPO): Direct Preference Optimization contrastive hardening across 4x NVIDIA RTX A6000 GPUs, penalizing scratchpad leakage (<think>), robotic refusals, raw SQL hallucinations, and schema deviations.
  • Zero-Reasoning Direct Response Invariant: By default, aiqlick/1.3 emits direct parseable AST JSON or tool calls without scratchpad <think> reasoning tokens or conversational fluff, achieving sub-100ms response times.
  • Held-out Ranking Accuracy: 86.46% preference accuracy on held-out validation pairs with a +0.1103 evaluation reward margin.

Serving with vLLM

vllm serve rzeraat/aiqlick-1.3 \
    --port 8000 \
    --max-model-len 4096 \
    --gpu-memory-utilization 0.8

Inference Example (OpenAI SDK)

from openai import OpenAI

client = OpenAI(base_url="http://localhost:8000/v1", api_key="EMPTY")

response = client.chat.completions.create(
    model="aiqlick/1.3",
    messages=[
        {"role": "system", "content": "You are aiqlick/1.3, an autonomous AI recruitment agent. You output direct JSON."},
        {"role": "user", "content": "Find senior backend developers in Stockholm with AWS and Go experience."}
    ],
    temperature=0.0
)
print(response.choices[0].message.content)
Downloads last month
635
Safetensors
Model size
4B params
Tensor type
BF16
ยท
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for rzeraat/aiqlick-1.3

Finetuned
Qwen/Qwen3.5-4B
Finetuned
(1,572)
this model