🧠 LFM2.5-350M-Thinking-0005
LFM2.5-350M-Thinking-0005 is the flagship lightweight 350-million-parameter Reasoning Model based on LiquidAI's Liquid Neural Network (LNN) architecture.
Fine-tuned on the Golden 10,709 Multi-Teacher Reasoning Mixture combining high-order multi-teacher logic (Qwen3.8-Max, GLM-5.2, Kimi-k3), DeepSeek-R1 Math, Russian R1 CoT, Bespoke Stratos, and Self-Awareness / Anti-Hallucination honesty calibration.
⚡ Highlights & Key Capabilities
- 🧬 Liquid Neural Network (LNN): Sub-linear memory scaling with extreme edge efficiency.
- 📐 Deep Chain-of-Thought: Structured
<think> ... </think>blocks before providing concise and correct final solutions. - 🇷🇺 Bilingual Russian & English Mastery: Fluent native Russian multi-step reasoning without translation noise.
- 🎯 Self-Awareness & Anti-Hallucination: Calibrated to detect unanswerable / hypothetical questions and answer honestly instead of hallucinating.
- 🚀 Extreme Edge Speed: Delivers 28–42+ tokens/sec on CPU and mobile devices.
📊 Dataset Mixture (Revision 0005)
- 🔗 Golden Dataset Repository:
KoarAI/LFM2.5-350M-Thinking-0005-Dataset - Total Training Samples: 10,709 Golden Samples
- Teachers: Qwen 3.8 Max, DeepSeek-R1, GLM-5.2, Kimi-k3, Bespoke Stratos.
📦 GGUF & Quantized Models
- 🔗 GGUF Repository: KoarAI/LFM2.5-350M-Thinking-0005-GGUF
LFM2.5-350M-Thinking-0005-f16.gguf(FP16 Baseline)LFM2.5-350M-Thinking-0005-q8_0.gguf(Q8_0 Quantized)