Notes

Uploaded in fp16 and Q8_0

• Q8_0 is the sweet‑spot for on‑device CPU inference
• fp16 is useful if you have enough GPU memory and want the absolute best performance

Temperature: The model was tuned for very deterministic output, so start with 0.1 – 0.2. Raise to ≈0.6 only if you need more creative or exploratory answers.

Description

LFM2.5‑1.2B‑Instruct – a 1.2B parameter, hybrid Liquid‑architecture model built for edge deployment. I’ve only uploaded the Instruct variant.

Key points:

• Hybrid design: 10 double‑gated LIV convolution blocks + 6 Grouped‑Query‑Attention (GQA) blocks give fast, low‑latency inference on CPUs, NPUs and mobile GPUs.
  
• Extended training: 28 T tokens of pre‑training plus multi‑stage reinforcement learning, delivering quality comparable to much larger models.
  
• Context length: 32 k tokens, 65 k‑vocab, multilingual (EN, AR, ZH, FR, DE, JA, KO, ES).
  
• Agentic ready: Supports function‑calling/tool use via the built‑in <|tool_call_start|> / <|tool_call_end|> token wrappers.
  
• Ideal use‑cases:
  
  • On‑device assistants, chatbots, and RAG pipelines
    
  • Structured JSON extraction / data‑wrangling
    
  • Lightweight code completion (FIM)
    
  • Edge AI in smartphones, laptops, vehicles, IoT devices
    

References

• Model Card (HF): https://huggingface.co/LiquidAI/LFM2.5-1.2B-Instruct
  
• Blog post (LFM2.5 launch): https://www.liquid.ai/blog/introducing-lfm2-5-the-next-generation-of-on-device-ai
  
• Documentation (Chat template & tool use): https://docs.liquid.ai/lfm/key-concepts/
  
• Inference guides:
  
  • Transformers → https://docs.liquid.ai/lfm/inference/transformers
    
  • llama.cpp → https://docs.liquid.ai/lfm/inference/llama-cpp