Show HN: Needle2: 14MB agentic LLM for phones, wearables, smart home and robots

What came through most clearly is that this only makes sense if you accept the product shape. People who treated the demo like a general language model got absurd outputs, including unrelated prompts mapping to door locks or thermostats. That did not really undermine the core claim. It clarified it. Needle 2 is being sold as a local intent-to-schema mapper where reliability depends on the confidence threshold, tight tool definitions, and task-specific fine-tuning. Several examples showed that wording of tool descriptions materially changes results, and the author repeatedly pointed users to the bundled data synthesis and fine-tuning pipeline rather than claiming strong zero-shot robustness. The useful criticism landed on evaluation, not on whether tiny models are interesting. People wanted abstention quality and false-positive rates measured across real command sets, not just anecdotal confidence scores from single prompts. That is the gap between a fun demo and something an integrator can trust in a lock, thermostat, or robot. At the same time, the mood was broadly positive because the target category feels real. A few commenters pointed out that the industry has overfocused on running ever-larger models on laptops and desktops, while there is obvious demand for a much smaller local action layer in voice interfaces, Home Assistant setups, cheap consumer devices, and lightweight robotics. Training cost also did not look like a serious objection at this scale. The more credible concern is product reliability, especially on unsupported or ambiguous commands.