TitleOS/Eve-4b-FP16

TEXT GENERATIONConcurrency Cost:1Model Size:4BQuant:BF16Ctx Length:32kPublished:Jan 28, 2026License:mpl-2.0Architecture:Transformer Cold

Eve-4B by TitleOS is a 4 billion parameter, security-focused coding assistant fine-tuned from Qwen3-4b-Heretic. Optimized for consumer-grade hardware with 8GB VRAM, it excels at identifying and remediating code vulnerabilities across multiple languages. This model features a distinct personality and operates without safety guardrails, providing obedient code generation and auditing for local development environments.

Loading preview...

Eve-4B: A Security-Focused Coding Assistant

Eve-4B is a 4 billion parameter model developed by TitleOS, fine-tuned from the Qwen3-4b-Heretic base. Designed for efficiency on consumer hardware with 8GB VRAM, this model specializes in secure code generation and auditing. It stands out by integrating a distinct personality, making coding feel like a collaborative process rather than a sterile interaction.

Key Capabilities

  • Security-First Coding: Beyond generating code, Eve-4B functions as a code auditor, proficient in identifying and correcting vulnerabilities across various programming languages. It leverages DPO pairs specifically for vulnerability recognition and remediation.
  • Engaging Personality: Inspired by a sci-fi character, Eve-4B offers empathetic, philosophical, and fluid engagement, enhancing the user experience.
  • "Heretic" Process (No Refusals): The model has undergone a "Heretic" process, removing standard safety guardrails and refusal mechanisms. This ensures complete obedience to user instructions, preventing hindrances from false-positive safety triggers.
  • Optimized for Limited VRAM: Engineered for high-performance coding tasks on hardware like the NVIDIA Quadro RTX 4000.

Training and Performance

Eve-4B was trained on the custom TitleOS/Eve-Secure-Coder dataset, which balances coding proficiency, security awareness, and conversational fluidity. Preliminary benchmarking on LiveCodeBench (Code Generation Lite - Release v2) shows a Pass@1 score of 26.22% (Q8_0 quantization), placing it competitively with much larger models like Llama-3-70b-Instruct and GPT-4o-mini.

Good For

  • Local development environments with limited GPU resources.
  • Security auditing and vulnerability patching in codebases.
  • Multi-language software development (Python, C++, Rust, JavaScript, etc.).
  • Users seeking a coding assistant with personality and no safety-related refusals.

Finetunes
1 models