Agent-Grade Intelligence, Compact by Design
Powering precise code generation and real-time tool execution, cisimcik-9b is the base model for developers who move with purpose — leveraging deep reasoning, multi-step agentic flows, and native tool-call support to accelerate every build.
The Builders
Two statistics students turning rigorous data into compact intelligence.
Çağrı Okan
Agent Reliability & Evaluation | AI Researcher
Istanbul-based statistics student, AI researcher, and full-stack developer. Focuses on agent reliability and evaluation, building intelligent systems at the intersection of statistics, software, and design. Hackathon veteran — built an AI warehouse inventory agent in 48 hours.
Yiğitcan Coşkun
Machine Learning Engineer | Statistics Student @YTU
3rd-year Statistics student at Yıldız Technical University focused on LLMs, ML, Deep Learning, and AI Agents. Bridging the curriculum gap by self-studying Data Structures & Algorithms and OOP — built MCP-based AI agent architectures and end-to-end Power BI pipelines for enterprise clients.
What Makes cisimcik-9b Different
Built on a rigorous data pipeline that prioritizes signal quality over volume.
Precision Fine-Tuning Engine
cisimcik-9b is trained on carefully curated data spanning coding, debugging, error handling, and agentic reasoning — across multiple programming languages and real production scenarios.
Data Monitoring Dashboard
A real-time dashboard tracks dataset health, coverage metrics, category distribution, and training signal quality — ensuring every sample added meaningfully improves the model.
Native Tool Call Support
Built-in XML-style tool calling supports no-tool, single-tool, multi-tool, and reasoning+tool flows out of the box — mirroring real production tool environments.
Structured Reasoning Blocks
Think blocks let the model reason through multi-step tasks, track constraints, and plan tool strategy before generating its final response.
Bilingual Intelligence
Native Turkish and English support with natural, non-robotic language. Turkish samples use real-world phrasing — never translated text in disguise.
Quality-First Dataset
High-signal samples only — covering success cases, edge cases, and failure recovery. Shallow, repetitive, or synthetic-sounding data is explicitly excluded.
Backed By
The people and organizations supporting cisimcik.