saidutta69/LFM2.5-350M-heretic

TEXT GENERATIONPricing:Input $0.04 / Cached $0.008 / Output $0.08Concurrent Unit Cost:1Model Size:0.35BQuant:BF16Context Size:32kPublished:Oct 9, 2026License:otherArchitecture:Transformer Featherless Exclusive Cold

LFM2.5-350M-heretic by saidutta69 is a 350 million parameter, decensored variant of LiquidAI's LFM2.5 model, featuring a hybrid convolution and attention architecture with a 32,768 token context length. It is specifically modified using the Heretic v1.4.0 tool to suppress refusal behavior via targeted weight edits, rather than fine-tuning, preserving its native Pythonic tool-call format and multilingual capabilities across nine languages. This model excels at data extraction, structured output, tool routing, and RAG on resource-constrained hardware, offering significantly faster decoding speeds than similarly sized dense LLMs.

Loading preview...

Overview

LFM2.5-350M-heretic is a 350 million parameter model developed by saidutta69, based on LiquidAI's LFM2.5. This variant is specifically decensored using the Heretic v1.4.0 tool, which employs "abliteration" (directional ablation) to suppress refusal behavior. Unlike traditional fine-tuning, abliteration directly edits specific weight directions responsible for refusals, leaving the base model's core agentic abilities, native Pythonic tool-call format, and multilingual coverage largely intact. This results in a significant reduction in refusals from 89/100 to 4/100, while maintaining the original model's performance characteristics.

Key Capabilities

  • Aggressively Decensored: Refusal behavior is drastically reduced, making it suitable for use cases requiring an uncensored assistant.
  • High Performance on Edge Devices: Decodes at 313 tok/s on AMD CPUs and 188 tok/s on Snapdragon Gen4 NPUs, significantly faster than dense LLMs of comparable size. It runs under 1 GB of memory, fitting on devices like phones, Raspberry Pi, or in a browser tab.
  • Multilingual Support: Capable of speaking nine languages (English, Arabic, Chinese, French, German, Japanese, Korean, Portuguese, Spanish).
  • Hybrid Architecture: Utilizes a unique hybrid convolution + attention architecture (Lfm2ForCausalLM) with 16 layers, optimized for decode-bandwidth rather than raw parameter count.
  • Tool Use & Structured Output: Retains Liquid's native Pythonic tool-call format and is effective for generating structured outputs.

Good For

  • Resource-Constrained Environments: Ideal for deployment on devices with limited RAM and processing power, where larger models are impractical.
  • Data Extraction & Structured Output: Excels at tasks requiring precise data extraction and generation of structured responses.
  • Tool Routing & Agentic Workflows: Maintains strong agentic abilities and tool-calling capabilities for complex automation.
  • RAG Applications: Suitable for Retrieval Augmented Generation (RAG) scenarios on edge hardware.
  • Developers Needing Uncensored Models: Provides an uncensored assistant without degrading the base model's core functionalities.

Finetunes
1 models