KoarAI/LFM2.5-350M-Thinking-0005

TEXT GENERATIONPricing:Input $0.04 / Cached $0.008 / Output $0.08Concurrent Unit Cost:1Model Size:0.35BQuant:BF16Context Size:32kPublished:Aug 31, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

KoarAI/LFM2.5-350M-Thinking-0005 is a 350-million-parameter reasoning model based on LiquidAI's Liquid Neural Network (LNN) architecture, featuring sub-linear memory scaling and extreme edge efficiency. It is fine-tuned on a multi-teacher reasoning mixture for deep chain-of-thought processing and bilingual Russian & English mastery. This model excels at structured reasoning, honest responses to unanswerable questions, and delivers high inference speeds on CPU and mobile devices.

Loading preview...

Overview

KoarAI/LFM2.5-350M-Thinking-0005 is a lightweight 350-million-parameter reasoning model developed by KoarAI. It is built upon LiquidAI's innovative Liquid Neural Network (LNN) architecture, which provides sub-linear memory scaling and is optimized for extreme edge efficiency. The model has a context length of 32768 tokens.

Key Capabilities

  • Liquid Neural Network (LNN): Offers efficient memory usage, making it suitable for edge devices.
  • Deep Chain-of-Thought: Utilizes structured <think> ... </think> blocks to produce concise and accurate solutions, enhancing its reasoning abilities.
  • Bilingual Mastery: Proficient in both Russian and English, enabling native multi-step reasoning without translation artifacts.
  • Self-Awareness & Anti-Hallucination: Calibrated to identify and honestly respond to unanswerable or hypothetical questions, reducing hallucination.
  • Extreme Edge Speed: Achieves high inference speeds of 28–42+ tokens/sec on CPU and mobile devices.

Training Details

The model was fine-tuned on the "Golden 10,709 Multi-Teacher Reasoning Mixture," which includes data from high-order multi-teacher logic models like Qwen3.8-Max, GLM-5.2, Kimi-k3, DeepSeek-R1 Math, Russian R1 CoT, Bespoke Stratos, and self-awareness calibration data. The dataset repository is available at KoarAI/LFM2.5-350M-Thinking-0005-Dataset.

Good For

  • Applications requiring efficient, on-device reasoning.
  • Tasks demanding structured, multi-step logical thinking.
  • Use cases where honest responses to ambiguous questions are critical.
  • Bilingual applications in Russian and English.

Finetunes
2 models