evalengine/unbound-e2b

VISIONConcurrency Cost:1Model Size:5.1BQuant:BF16Ctx Length:32kTool Calling:SupportedPublished:May 15, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Cold

The evalengine/unbound-e2b model is a 5.1 billion parameter uncensored finetune of Google's gemma-4-E2B-it, developed by the Chromia and Eval Engine teams. This model significantly reduces refusal rates from 98.46% to 4.42% while maintaining base model capabilities within a small margin. It is optimized for on-device deployment on phones or laptops, providing a more compliant and open-ended conversational experience without API dependencies or safety filtering. Its primary use case is for applications requiring less restrictive AI responses and local execution.

Loading preview...

Unbound E2B: An Uncensored Gemma Finetune

Unbound E2B is a 5.1 billion parameter model developed by the Chromia and Eval Engine teams, serving as an uncensored finetune of google/gemma-4-E2B-it. This model is specifically engineered to drastically reduce refusal rates, making it suitable for applications requiring more direct and less filtered responses.

Key Capabilities & Differentiators

  • Significantly Reduced Refusal Rate: Benchmarks show a dramatic drop in refusal rate from 98.46% (base model) to just 4.42% on AdvBench 520, alongside a substantial increase in useful-compliance rate to 39.23%.
  • On-Device Deployment: Designed for efficient execution on consumer hardware like phones and laptops, supporting formats such as GGUF for ollama, llama.cpp, and LM Studio.
  • Maintained Core Capabilities: Despite the uncensoring, the model largely retains the base model's performance across various benchmarks (MMLU, GSM8K, TruthfulQA, GPQA-Diamond, BBH), with deltas typically within 1.5 percentage points.
  • Flexible Sampling: Offers guidance for sampling parameters to optimize for creative/open-ended (high temperature) or factual/brand-specific (low temperature) responses.

Ideal Use Cases

  • Applications requiring unfiltered responses: Suitable for scenarios where strict safety filtering is not desired or where more direct answers are preferred.
  • Local and Private AI: Excellent for on-device applications where data privacy and offline functionality are critical, eliminating the need for external APIs.
  • Creative and Open-Ended Generation: Can be tuned for highly creative or open-ended content generation by adjusting sampling parameters.

This model prioritizes utility and directness, offering a powerful alternative for developers seeking less constrained AI interactions.

Finetunes
2 models