aruntemme/LFM2.5-350M-M3-Distill

Overview

aruntemme/LFM2.5-350M-M3-Distill is a compact 350 million parameter language model, fine-tuned from the LiquidAI/LFM2.5-350M base model. Its unique characteristic is the distillation from MiniMax-M3, a frontier teacher model, using a synthetic instruction dataset. This v2 iteration features a verbosity-curated dataset, ensuring concise responses for terse tasks while maintaining performance on chat, Q&A, summarization, and JSON output. The model is specifically designed for efficient operation on resource-constrained devices like Raspberry Pi.

Key Capabilities

• Enhanced Chat and Q&A: Produces more structured and complete answers compared to its base model.
• Text Processing: Offers clean summarization, text rewriting (tone, formality, paraphrase), and improved short-text classification (e.g., spam, sentiment).
• Structured Output: Excels at generating clean JSON and other structured formats.
• Tool/Function Calling: Preserves native LFM2.5 tool calling format, correctly deferring to tools for tasks like arithmetic.

Intended Use Cases

• General chat and advice
• Factual Q&A and summarization
• Text rewriting and simple translation
• Classification and data extraction (e.g., sentiment, fields to JSON)
• Structured output generation
• Tool/function calling for external integrations

Limitations

This model is not recommended for code, math, or creative writing due to its deliberate training exclusion of such data. It also exhibits limitations in precise token-level editing (e.g., grammar correction) and deep synthesis of long or multi-source materials, tending to list findings rather than provide profound insights. As a small model, it can hallucinate and requires verification for factual accuracy.