Show HN: Needle2: 14MB agentic LLM for phones, wearables, smart home and robots https://ift.tt/wS5y2TL

Show HN: Needle2: 14MB agentic LLM for phones, wearables, smart home and robots Hey HN, Henry from Cactus here! We previously released Cactus Needle, a 14MB agentic LLM for tool call, device use, and structured extraction for phones, wearables, smart homes, small robots and microcontrollers. We got really great feedback here, and have now incorporated the suggestions to release Needle 2. The whole model is a single 14MB binary that runs a full session in 28MB of RAM; 45m parameters at 2bit compression. Needle hits 500 tokens/sec decode speed on a Raspberry Pi 5, sits between 400-1,500 tokens/sec on VR devices like Meta Quest 3S and Apple Vision Pro, and ranges 300-700 on sub-$200 phones such as the Samsung A-Series. On the tool call and mobile device use benchmarks, Needle 2 trades wins with closest small models like LFM2.5 230M and Apple Foundation Model, at 5x to 70x smaller, both at f16 vs Needle 2 at 2bit. Needle is based on Simple Attention Networks from our paper ( https://ift.tt/Z7O9KUj ). Edge AI has lately meant Macs and PCs, but that is just 1.5 billion of over 21 billion connected IoT devices in the world today, and in emerging markets most phones ship under $200, no NPU, cheap GPUs. These include budget phones, Raspberry Pis, microcontrollers, wearables, small robots like Reachy Mini, and connected home devices. A conventional transformer of Needle's width and depth spends 164 MFLOPs per token, and even one squeezed down to Needle's parameter count spends 87, Needle spends 70. Even on a high-end phone, an always-on assistant lives inside a power budget; every MFLOP is milliwatt-hours, and Needle spends 7x to 85x fewer of them per token than the smallest performant LLMs. More about the architecture in the link. When we structure intelligence for consumer devices as functions with typed parameters, the only hard part is mapping a messy sentence onto them; which function, with which values. Our research found that when framed that way, the problem needs no world knowledge and no open-ended prose, which is why 45M parameters suffice. Needle 2 expands to structured extraction where the schema can be passed in-place of tools and the model returns structured output. You can use Needle as a text-classification model with an enum field, as a summarization model by providing a schema that extracts key fields, everything but free-range decode. Every product has its own tool vocabulary and fine-tuning needle helps it achieve frontier-level performance on custom tasks, so using the python package ( https://ift.tt/xGTEdZD ), Needle can be fine-tuned Needle on a Mac/PC in minutes to a few hours, with automated data-generation pipeline, just pass a couple samples. Nonetheless, every response carries a learned confidence score based our Cactus Hybrid technique. If above your threshold, act, below it, escalate to the cloud or bigger model. Combining Needle 2 with a private DeepSeek-v4-Flash deployment works particularly well for enterprise-level tasks at barely any cost, we can help with this setup. We have put a lot of thoughts into Needle 2 but might still be missing quite a lot, please use the playground in the provided link to test Needle and share your thoughts, always appreciated! https://ift.tt/c4tQqnD August 10, 2026 at 10:52PM

Show HN: 100% native Swift harness (NOT Electron) https://ift.tt/EXfWiPV

Show HN: 100% native Swift harness (NOT Electron) hi everybody, I’ve been working on this harness that is all native Swift for macOS. It’s fully featured with every feature i could find in cline, codex, and claude code. There is all swift command line too It’s taking me a long, long time, a few months to build this, and it’s fully open source Apache 2 licensed. Hoping people fork this and make many branches I’m looking for people who want to work on something that has much lower memory usage and much more potential to be just more efficient. Every other single harness has is written in using electron, so this harness is really unique in this way. It’s connected to trusted router so that it has end-to-end encryption confidential mode as well as access to hundreds of models, any model in the world. It works great with Deepseek. it’s great for swift development ios development. and for people who love Linux, I also made a compatibility layer that lets you compile SwiftUI into native Linux so that you either can new or QT actually so that you can just compile all the code directly and make it run on Linux https://ift.tt/5w8n1OG August 10, 2026 at 06:56PM

Show HN: Run your GitHub Actions locally or self-hosted in isolated microvms https://ift.tt/d4i6syT

Show HN: Run your GitHub Actions locally or self-hosted in isolated microvms Hey HN. I've been frustrated with Github Actions' reliability, and inability to run workflows locally. Preloop is a Rust reimplementation of the GitHub Actions, both the runner(tracking the latest version) and the control plane that runs in hardware-isolated microvms on MacOS/Linux/Windows(we use smolvm project that uses the libkrun vmm).The microvms starts in <400ms from a packed artifact, and cow-clones for each job are instant. The control plane speaks the exact official runner protocol, so the unmodified actions/runner binary can register, poll, execute, and report against it. Your existing workflow files run unchanged, but on your own hardware with zero GitHub-hosted minutes. Locally, you can also run uncomitted changes so you dont have to create a commit to run ci. If you are working solo or in a small team and have fast local hardware, you can use something like DHH's gh-signoff extension and merge directly so long as certain checks pass. We also support features such as pause-on-failure where upon failure, the vm is paused, opens up a shell, and allows you to fix the errors, then you can retry without rerunning all the jobs that worked(new step/job runs in a new fresh microvm to verify the changes) or you can attach directly to a running vm and see what steps each jobs are doing in real-time. Preloop also implements the Debug Adapter Protocol so you or your agent can drive your CI locally(with the built-in client) or you can use a DAP client like Vscode. You can also submit your changes to the server for CI to run and then after it can open a draft PR for you, with the updated checks, so we dont only rely on Webhooks from Github. I've invested heavily in ensuring conformance with the official runner protocol so I do different kinds of mitm proxying to make sure not only the request/response bodies match but also down to the job/step level conclusions/logs/annotations etc. Effectively, the official runner should work with our control plane. Act, Forgejo are all great, but they dont faithfully implement the official runner protocol so in practice some workflows dont quite work out of the box, and they also run the jobs in docker containers. Forgejo has the closest compatibility and is probably still a good option if you are fully migrating off github. With Act, there's a few mostly server-side features it doesnt implement(concurrency, some matrix features, oidc etc). forgejo has the closest compatibility, and is a good option if you are fully migrating off Github. Agent-ci is also a great option that uses the official runner image and implements the official protocol, but it is designed to run locally, runs in containers, and doesnt support some features we do(concurrency, newer V2 cache, Github token etc). Preloop is designed to also be run as a peristent server so we implement more submit updates to the checks api. Lot more stuff coming, but i figured this is a good point to share with the community incase it might be interesting to some folks. Feel free to try it out. Website: https://preloop.dev Github: https://github.com/preloopdev/preloop https://preloop.dev August 10, 2026 at 01:25AM

Show HN: Inertia – Not your Grandfathers animation editor https://ift.tt/0PbBOYa

Show HN: Inertia – Not your Grandfathers animation editor https://ift.tt/K37HyMS August 9, 2026 at 03:57AM

Show HN: A Discord bot running in a 4MiB microVM https://ift.tt/rx8diA3

Show HN: A Discord bot running in a 4MiB microVM https://ift.tt/GquysbB August 8, 2026 at 11:04PM

Show HN: Algebruh - Cross-check arithmetic claims with Z3, cvc5, and Lean https://ift.tt/ytLl4vO

Show HN: Algebruh - Cross-check arithmetic claims with Z3, cvc5, and Lean https://ift.tt/fcNwrIH August 8, 2026 at 11:34PM

Show HN: Check if any of the $656M in unclaimed royalties at The MLC is yours https://ift.tt/qYFMHKa

Show HN: Check if any of the $656M in unclaimed royalties at The MLC is yours https://pub.doub.ly/ August 7, 2026 at 11:05PM

Show HN: Zaivern Code – a Rust cockpit for parallel AI coding agents https://ift.tt/gC24wh3

Show HN: Zaivern Code – a Rust cockpit for parallel AI coding agents https://ift.tt/JdxwrWs August 8, 2026 at 12:53AM

Show HN: Mousecrack – Teaching an LSTM to move a mouse like a human https://ift.tt/t1VPjba

Show HN: Mousecrack – Teaching an LSTM to move a mouse like a human https://ift.tt/0Tfzg8v August 7, 2026 at 10:23PM

Show HN: ARF – a record format for AI evaluation runs, with reproducible digests https://ift.tt/eCOY5Hy

Show HN: ARF – a record format for AI evaluation runs, with reproducible digests https://www.korvo.xyz/arf August 7, 2026 at 04:55AM

Show HN: AI Tutoring with Visual Grounding https://ift.tt/pXdYANx

Show HN: AI Tutoring with Visual Grounding I've always felt that AI tutoring as we see it now is going in the wrong direction. I made Knowable as a way to see if it's possible to have a real AI tutor. Because of hardware limitations it can only be used on Macbooks 2023+. Let me know your thoughts! https://useknowable.ai/ August 7, 2026 at 05:34AM

Show HN: Learn System Design, one campaign at a time https://ift.tt/ScTMFku

Show HN: Learn System Design, one campaign at a time https://scalequest.io/ August 7, 2026 at 03:25AM

Show HN: Pokémon Emerald Ported to Raspberry Pi Pico 2 https://ift.tt/LxUSFWg

Show HN: Pokémon Emerald Ported to Raspberry Pi Pico 2 Pokémon Emerald ported to the RP2350 microcontroller. No emulator, 60 fps HDMI output. Recompiled from ARMv4T to Cortex-M33 and the Game Boy Advance's video hardware is reimplemented in software on the second core. https://ift.tt/4Xkgre5 August 7, 2026 at 03:19AM

Show HN: LiminalML – Study ML or SWE at interview depth, grounded in your resume https://ift.tt/TPu2D7a

Show HN: LiminalML – Study ML or SWE at interview depth, grounded in your resume https://liminalml.com August 5, 2026 at 11:09PM

Show HN: Maple-Preview – ternary 20B MoE running at 120 tok/s on a iPhone https://ift.tt/jUEVPz5

Show HN: Maple-Preview – ternary 20B MoE running at 120 tok/s on a iPhone https://ift.tt/QJCH81n August 5, 2026 at 01:14AM

Show HN: My tool scanned 256 AI-built apps and most had exposed credentials https://ift.tt/Y3xCTOG

Show HN: My tool scanned 256 AI-built apps and most had exposed credentials All of a sudden everybody wants to build with AI. People in hell want ice water. Built something with AI and don't know if it's secure? Try necktochoke. https://ift.tt/45A2bvw August 5, 2026 at 01:42AM

Show HN: TormentNexus – Local-first Go control plane with persistent memory https://ift.tt/ctsfHVN

Show HN: TormentNexus – Local-first Go control plane with persistent memory https://tormentnexus.site August 5, 2026 at 12:47AM

Show HN: Spellfolio – my stock market game with news events and inside info https://ift.tt/gzYGOsU

Show HN: Spellfolio – my stock market game with news events and inside info https://ift.tt/4aZRHLn August 4, 2026 at 04:08AM

Show HN: Freqcast, an Android radio player that finds streams from a website URL https://ift.tt/pUQ0gZB

Show HN: Freqcast, an Android radio player that finds streams from a website URL Hi HN! Not every internet radio station is listed in a directory like Radio Browser. For a lot of them, the only way to actually listen is to dig the stream URL (.mp3/.aac/.m3u8) out of the station's website by hand — and that's before you even find a player that isn't full of ads or locked to its own catalog. Freqcast lets you paste the station's website instead of the stream URL directly. It first checks the Radio Browser catalog, then scans the website itself, discovers a playable stream, verifies that it works, and adds it to your library. In many cases, you never have to hunt for the URL yourself. No ads, no tracking, no account — everything lives locally in the app. Your station list can be exported as JSON, and imported from JSON, OPML, M3U, or PLS. No Play Store listing — grab the APK from Releases. F-Droid submission is in progress. Would genuinely love feedback! https://ift.tt/JlFgTdY August 4, 2026 at 03:05AM

Show HN: Golars (Go Equivalent of Polars) https://ift.tt/JACc1dh

Show HN: Golars (Go Equivalent of Polars) https://ift.tt/tXwhgnA August 4, 2026 at 01:30AM