Cactus releases Needle 2, a 14MB AI model built for phones, wearables, and IoT devices

·1 views

Cactus has launched Needle 2, a 45-million-parameter language model compressed to just 14MB that runs on resource-constrained devices including budget smartphones, Raspberry Pis, wearables, and small robots. The model operates within 28MB of RAM and achieves 300–500 tokens per second on sub-$200 Android phones and single-board computers, targeting the roughly 21 billion connected IoT devices worldwide that lack dedicated AI hardware. Needle 2 is designed specifically for structured tasks such as tool calling, device control, and data extraction, with the developers arguing that these functions require no broad world knowledge, making a small parameter count sufficient. On relevant benchmarks, it competes with models five to seventy times larger, including Apple's Foundation Model and LFM2.5 230M, while consuming significantly fewer compute resources per token. The model supports fine-tuning on a standard laptop in minutes to a few hours and includes a confidence-scoring mechanism that can escalate low-certainty requests to a larger cloud-based model.

Read the full story at Hacker News

This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)

Rallylens App Uses AI Video Analysis to Give Tennis Players Coach-Like Feedback

A student tennis player has developed Rallylens, a web app that uses AI-powered video analysis to help players improve their technique without a coach present. The tool breaks uploaded tennis videos into frames, classifies stroke types such as forehands, backhands, and serves, and compares them against biomechanical benchmarks with an 85% accuracy rate. When flaws are detected, the system generates specific, actionable feedback — for example, recommending precise adjustments to shoulder tilt or wrist rotation. The app has notable limitations, including sensitivity to poor video quality, difficulty with unconventional playing styles, and an inability to account for psychological factors affecting performance. Its creators position Rallylens not as a coach replacement but as a supplementary tool to make quality technique feedback more accessible, especially for players in remote areas or with limited coaching budgets.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

How One Developer Fixed Date Localization in React Native Using i18next and Intl API

A developer building a React Native app found that after integrating i18next for text translation, date formats like '7月24日(金)' continued to display in Japanese regardless of the selected language. The root cause was that dates formatted directly via JavaScript's Date object bypassed the translation pipeline entirely. To fix this, the developer unified language detection by storing user preferences in SecureStore and routing both i18next and Intl.DateTimeFormat through the same language setting. A custom formatScheduleDate function was built using Intl.DateTimeFormat to render dates correctly in either Japanese or English based on the resolved language. The developer noted that true multi-language support must extend beyond JSON translation strings to cover all dynamically generated content, including dates, times, and error messages.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

How a Serverless AWS Lambda Container Segments and Straightens Card Photos Using AI

A developer has built an image segmentation pipeline on AWS Lambda that takes a crooked, background-filled photo of an AWS Builder Card and returns a clean, cropped, straightened version. The system uses two Lambda functions: one to classify whether the uploaded image contains a valid card, and a second called image-processor to isolate and clean it. The segmentation work is handled by the BiRefNet-General-Lite model, loaded via the rembg library and running on CPU through ONNX Runtime, with no PyTorch required. Memory management proved a key challenge, as the full BiRefNet-General model caused out-of-memory errors even on AWS Lambda's largest available configuration. The lighter Lite variant was ultimately adopted to stay within Lambda's memory limits, with costs and vCPU allocation tied directly to the memory size chosen.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

TRC-20, BEP-20, Arbitrum or TON: Choosing the Right Chain for USDT Payouts

Developers building automated payout systems for platforms like freelance marketplaces or AI agent networks increasingly rely on USDT stablecoins across multiple blockchains instead of traditional bank transfers. TRC-20 on the Tron network remains popular for high-volume, low-value payouts due to fees as low as $0.50–$1.00 and near-universal exchange support. BEP-20 on BNB Smart Chain offers EVM compatibility and Solidity smart contract support, making it suited for treasury and agent-to-agent payment logic. Arbitrum, an Ethereum Layer 2, provides the lowest fees for complex operations at $0.10–$0.30 per transfer while inheriting Ethereum's security, though it requires users to understand bridging. Each chain involves distinct trade-offs around fees, developer tooling, exchange support, and user familiarity, meaning the best choice depends on the specific use case and audience.

0 comments Read more at DEV Community