saidutta69/LFM2.5-1.2B-Instruct-heretic

TEXT GENERATIONPricing:Input $0.04 / Cached $0.008 / Output $0.08Concurrent Unit Cost:1Model Size:1.2BQuant:BF16Context Size:32kPublished:Oct 9, 2026License:otherArchitecture:Transformer Featherless Exclusive Cold

saidutta69/LFM2.5-1.2B-Instruct-heretic is a 1.17 billion parameter instruction-tuned causal language model, a decensored variant of LiquidAI/LFM2.5-1.2B-Instruct. It utilizes a hybrid conv+attention architecture with a 32K context length, optimized for efficient local deployment across various devices. This model significantly reduces refusal behavior from 98/100 to 2/100 through targeted weight edits (abliteration) rather than fine-tuning, preserving its original instruction-following and tool-call capabilities. It is designed for local agents, on-device assistants, RAG, and structured extraction where a small, fast, and genuinely agentic model is required without GPU dependency.

Loading preview...

LFM2.5-1.2B-Instruct-heretic Overview

This model is a 1.17 billion parameter instruction-tuned language model, a "decensored" variant of LiquidAI/LFM2.5-1.2B-Instruct. It was created using the Heretic v1.4.0 tool, which employs directional ablation ("abliteration") to suppress refusal behavior. Unlike traditional fine-tuning, abliteration directly edits specific weight directions responsible for refusal, leaving the base model's core agentic abilities, native Pythonic tool-call format, instruction-following, and multilingual coverage largely intact. This results in a significant reduction in refusals from 98/100 to 2/100 on harmful evaluation sets, with a moderate KL divergence of 0.0657.

Key Capabilities

  • Reduced Refusal Behavior: Achieves a refusal rate of 2/100, down from 98/100 in the base model, enabling compliance with a broader range of requests.
  • Efficient Local Deployment: Designed to run effectively on devices with limited resources, including phones, Raspberry Pi, and laptops, without requiring a dedicated GPU.
  • Strong Instruction Following: Maintains the base model's high instruction-following capabilities (86.2 on IFEval) and agentic performance (49.1 on BFCLv3).
  • Hybrid Architecture: Utilizes a unique hybrid convolution + attention architecture, making it decode-bandwidth-bound and performant with fast RAM.
  • Multilingual Support: Covers English, Arabic, Chinese, French, German, Japanese, Korean, and Spanish.
  • Pythonic Tool-Call Format: Retains the base model's native support for Pythonic tool calls, facilitating integration with agentic workflows.

Good For

  • Local Agents and On-Device Assistants: Ideal for applications requiring a small, fast, and capable model that can run directly on user devices.
  • RAG (Retrieval-Augmented Generation): Suitable for tasks needing structured extraction and information retrieval where a compact model is advantageous.
  • Developers Needing Unfiltered Responses: Specifically targets developers who require instruction-following without inherent refusal mechanisms, while understanding the responsibility for deployment.
  • Resource-Constrained Environments: Excellent choice for scenarios where GPU access is limited or power efficiency is critical.

Finetunes
1 models