Tiny models getting genuinely useful for edge and mobile
Three separate threads landed on the same day about small and tiny models doing more than expected. Needle2 is a 14MB agentic LLM designed for phones, wearables, and robots. LiquidAI's LFM2.5 at 2.6B is reportedly competitive with models four times its size. And the H3-metal project is bringing MiniMax-H3 inference to Apple Silicon, with commenters running it on M5 Pro MacBook Pros via ComfyUI.
The through-line is that 'small model' no longer means 'toy.' The LiquidAI approach in particular is notable because they target reliable operation of tiny models rather than just scaling down a larger architecture. The Needle2 demo got mixed reception on accuracy, but the concept of a fully local, function-calling-capable model at 14MB is a genuine engineering milestone.
The key tension in the comments: people are excited about what's possible but honest about current limitations. Needle2's web demo produced some clearly wrong results. The question being debated is how much knowledge can actually be compressed into very small models, and whether fine-tuning on narrow domains closes the gap enough for real production use.
So what?
If you're building apps that need any AI capability on-device, this space is moving fast enough to revisit every six months. For consumer apps especially, local inference means no API costs, no latency, and no privacy concerns to explain to users. The fine-tuning angle is the most actionable part: a tiny model fine-tuned on your specific domain may outperform a general model twenty times its size.
Read these
Show HN: Needle2: 14MB agentic LLM for phones, wearables, smart home and robots
LFM2.5 2.6B model competitive with 4x larger models
Meta Muse Glimmer – open weights 30B local coding model
H3-metal – Native MiniMax-H3 inference for Apple Silicon