Needle 2 is an agentic LLM designed for tool use, device operation, and structured extraction. The model consists of a single 14MB binary and requires 28MB of RAM to operate. It has 45M parameters and utilizes 2-bit compression.
Inference speeds are exceptionally fast. It reaches 500 tokens/sec on a Raspberry Pi 5 and between 400 to 1,500 tokens/sec on Meta Quest 3S and Apple Vision Pro. According to Hacker News, it also operates at 300 to 700 tokens/sec on low-cost smartphones priced under $200.
Regarding performance, Cactus claims that Needle 2 competes in benchmarks with other small-scale models such as LFM2.5 230M and the Apple Foundation Model, while being 5 to 70 times smaller. It is based on Simple Attention Networks, which the company presented in a paper arXiv:2607.18363.
Power efficiency is also a primary focus. The computation per token is 7 to 85 times lower than existing smallest-class LLMs. It is intended to run on inexpensive hardware without an NPU, aligning with the global proliferation of IoT devices.
Additionally, structured extraction capabilities have been expanded. By passing a schema as a tool, users can obtain structured outputs for tasks such as text classification and summarization. According to Hacker News, fine-tuning can be completed in a few minutes to a few hours using the Python package github.com/cactus-compute/needle.
Source: Show HN: Needle2: 14MB agentic LLM for phones, wearables, smart home and robots（HN 537pt・185コメント） (HN Search (backfill), 2026-08-11)