KoarAI/LFM2.5-350M-Thinking

TEXT GENERATIONPricing:Input $0.04 / Cached $0.008 / Output $0.08Concurrent Unit Cost:1Model Size:0.35BQuant:BF16Context Size:32kPublished:Aug 27, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

KoarAI/LFM2.5-350M-Thinking is an ultra-compact 350 million parameter language model from KoarAI, built on the Liquid Foundation Model (LFM2.5-350M) architecture. This model is specifically fine-tuned for native Chain-of-Thought (CoT) reasoning, demonstrating strong multi-step logic and mathematical deduction. It excels at structured problem-solving within dedicated blocks, making it suitable for tasks requiring explicit reasoning processes.

Loading preview...

KoarAI/LFM2.5-350M-Thinking: Compact Reasoning Model

This model, designated as Code: 0002, is an ultra-compact, high-efficiency language model developed by KoarAI. It features native Chain-of-Thought (CoT) reasoning capabilities and is built upon the Liquid Foundation Model architecture (LiquidAI/LFM2.5-350M).

Key Capabilities & Training

  • Native Reasoning: The model is designed to reason internally before generating a final response, utilizing <think> ... </think> blocks for explicit multi-step logic and mathematical deduction.
  • Compact Size: Despite having only 350 Million parameters, it demonstrates strong structured problem-solving abilities.
  • Full Parameter Fine-Tuning: It underwent 100% full parameter fine-tuning over 9 epochs, integrating an expanded multi-teacher dataset.
  • Distilled Reasoning Traces: Training data included distilled reasoning traces from frontier models such as Qwen 3.8 Max, GLM 5.2, Kimi K3, DeepSeek-V4-Pro 0813 Agentic, and specialized datasets like MMLU-Pro and AIME 2026 Mathematics.

Use Cases

This model is particularly well-suited for applications requiring:

  • Explicit Reasoning: Tasks where the intermediate steps of problem-solving are important or need to be auditable.
  • Mathematical Deduction: Solving arithmetic and more complex mathematical problems with step-by-step logic.
  • Structured Problem-Solving: Scenarios demanding a structured approach to arrive at a solution.

Quantized GGUF versions are available for llama.cpp, Ollama, and LM Studio at KoarAI/LFM2.5-350M-Thinking-GGUF.