FlameF0X/LFM2.5-1.2B-Distilled-Claude-4.6
FlameF0X/LFM2.5-1.2B-Distilled-Claude-4.6 is a 1.2 billion parameter language model distilled from Claude Sonnet 4.6 Adaptive Thinking into the LFM2.5-1.2B-Thinking architecture using LoRA. This model is specifically trained on private data featuring agentic and pure reasoning 'think blocks' and consecutive user messages, aiming to replicate Claude's reasoning patterns. It is designed for tasks requiring structured thought processes and agentic behavior, as evidenced by its training on assistant messages with tool calls.
Loading preview...
LFM2.5-1.2B-Distilled-Claude-4.6 (Liquid Claude)
This model, developed by FlameF0X, is a 1.2 billion parameter language model distilled from Claude Sonnet 4.6 Adaptive Thinking (the claude.ai version, not the API) into the LFM2.5-1.2B-Thinking architecture using LoRA. Its training data is private and specifically designed to imbue the model with Claude's reasoning capabilities.
Key Training Details & Capabilities
- Distillation Focus: The model's core differentiator is its distillation from a highly capable Claude version, aiming to transfer advanced reasoning and agentic thinking into a smaller model.
- Think Block Patterns: Training data includes 8473 agentic 'think blocks' (Action/Observation) and 3005 pure reasoning 'think blocks', totaling over 15 million approximate tokens.
- Agentic Behavior: Approximately 18.3% of assistant messages in the training data include agentic tool calls within their thought processes, indicating a focus on structured problem-solving.
- Context Length: Supports a context length of 32768 tokens.
Performance
On the Intel/low bit open LLM leaderboard, LFM2.5-1.2B-Distilled-Claude-4.6 achieves an average score of 46.76, with specific scores of 39.51 on HellaSwag, 31.99 on MMLU, and 68.77 on Piqa. This shows an improvement over its base, FlameFOX/LFM2.5-1.2B-Thinking-CodeX, particularly in MMLU scores.