DavidAU/LFM2.5-1.2B-MEGABRAIN-Thinking-Claude-Polaris-Deepseek-GLM

TEXT GENERATIONConcurrent Unit Cost:1Model Size:1.2BQuant:BF16Context Size:32kPublished:Feb 6, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

DavidAU/LFM2.5-1.2B-MEGABRAIN-Thinking-Claude-Polaris-Deepseek-GLM is a 1.2 billion parameter language model, fine-tuned by DavidAU, that completely replaces the original reasoning capabilities of the LFM2.5 base model. This model is optimized for deep and detailed reasoning, providing compact yet thorough outputs. It features a 128k context window and stable reasoning across a wide temperature range, making it suitable for tasks requiring precise and consistent logical processing.

Loading preview...

Model Overview

DavidAU/LFM2.5-1.2B-MEGABRAIN-Thinking-Claude-Polaris-Deepseek-GLM is a 1.2 billion parameter language model developed by DavidAU, built upon the LFM2.5 architecture. This model has undergone a comprehensive fine-tuning process using multiple datasets via Unsloth, specifically targeting and replacing the base model's reasoning and thinking capabilities. The result is a model designed to produce compact, yet highly detailed and precise reasoning outputs.

Key Features & Enhancements

  • Enhanced Reasoning: The core differentiator is its completely re-engineered reasoning engine, leading to more detailed and to-the-point logical processing.
  • Extended Context Window: Supports a substantial 128k token context length, allowing for processing of longer inputs and maintaining coherence over extended conversations or documents.
  • Temperature Stable Reasoning: The model's reasoning performance remains consistent across a broad temperature range (0.1 to 2.5), offering flexibility in output creativity without compromising logical integrity.
  • Improved Benchmarks: Compared to the "Normal LFM2.5" base, this fine-tune shows notable improvements across several benchmarks, including ARC-Challenge (0.408 vs 0.365), ARC-Easy (0.589 vs 0.426), BoolQ (0.771 vs 0.717), and Hellaswag (0.572 vs 0.486).

Optimal Usage

For best performance, the model's creators strongly recommend using q5, q6, q8, or 16-bit precision, or Imatrix IQ3_M minimum. A repetition penalty of 1.05 to 1.1 is suggested. For chat, roleplay, or smoother operation, users are advised to set a "Smoothing_factor" to 1.5 in interfaces like KoboldCpp, oobabooga/text-generation-webui, or Silly Tavern. Detailed guidance on maximizing model performance and advanced settings is available in the provided Maximizing Model Performance guide.