frontpage.
newsnewestaskshowjobs

Made with ♥ by @iamnishanth

Open Source @Github

fp.

Open in hackernews

Show HN: Needle: We Distilled Gemini Tool Calling into a 26M Model

https://github.com/cactus-compute/needle
45•HenryNdubuaku•1h ago
Hey HN, Henry here from Cactus. We open-sourced Needle, a 26M parameter function-calling (tool use) model. It runs at 6000 tok/s prefill and 1200 tok/s decode on consumer devices.

We were always frustrated by the little effort made towards building agentic models that run on budget phones, so we conducted investigations that led to an observation: agentic experiences are built upon tool calling, and massive models are overkill for it. Tool calling is fundamentally retrieval-and-assembly (match query to tool name, extract argument values, emit JSON), not reasoning. Cross-attention is the right primitive for this, and FFN parameters are wasted at this scale.

Simple Attention Networks: the entire model is just attention and gating, no MLPs anywhere. Needle is an experimental run for single-shot function calling for consumer devices (phones, watches, glasses...).

Training: - Pretrained on 200B tokens across 16 TPU v6e (27 hours) - Post-trained on 2B tokens of synthesized function-calling data (45 minutes) - Dataset synthesized via Gemini with 15 tool categories (timers, messaging, navigation, smart home, etc.)

You can test it right now and finetune on your Mac/PC: https://github.com/cactus-compute/needle

The full writeup on the architecture is here: https://github.com/cactus-compute/needle/blob/main/docs/simp...

We found that the "no FFN" finding generalizes beyond function calling to any task where the model has access to external structured knowledge (RAG, tool use, retrieval-augmented generation). The model doesn't need to memorize facts in FFN weights if the facts are provided in the input. Experimental results to published.

While it beats FunctionGemma-270M, Qwen-0.6B, Granite-350M, LFM2.5-350M on single-shot function calling, those models have more scope/capacity and excel in conversational settings. We encourage you to test on your own tools via the playground and finetune accordingly.

This is part of our broader work on Cactus (https://github.com/cactus-compute/cactus), an inference engine built from scratch for mobile, wearables and custom hardware. We wrote about Cactus here previously: https://news.ycombinator.com/item?id=44524544

Everything is MIT licensed. Weights: https://huggingface.co/Cactus-Compute/needle GitHub: https://github.com/cactus-compute/needle

Comments

simonw•40m ago
Looks like you need to open up access to https://huggingface.co/Cactus-Compute/datasets/needle-tokeni... - I get this error when trying to run the steps in your README:

> Repository Not Found for url: http s://huggingface.co/api/datasets/Cactus-Compute/needle-tokenizer/revision/main.

HenryNdubuaku•33m ago
Fixed now, apologies!
simonw•10m ago
Thanks, works now: https://gisthost.github.io/?4ff455792651fe755265b467800f47f3
ilaksh•25m ago
Hmm.. this might make it feasible to build something like a command line program where you can optionally just specify the arguments in natural language. Although I know people will object to including an extra 14 MB and the computation for "parsing" and it could be pretty bad if everyone started doing that.

But it's really interesting to me that that may be possible now. You can include a fine-tuned model that understands how to use your program.

E.g. `> toolcli what can you do` runs `toolcli --help summary`, `toolcli add tom to teamfutz group` = `toolcli --gadd teamfutz tom`

HenryNdubuaku•21m ago
So Needle is trained for INT4, what you see in the playground is INT4, only 14MB, same challenge though.
ilaksh•16m ago
Oh gotcha. Fixed my comment.
cmrdporcupine•9m ago
This is very cool I'm going to try to carve out some time to try building this into my MOO system ( https://codeberg.org/timbran/moor / https://timbran.org/moor.html ) as alternative command parser front end.
simonw•7m ago
Suggestion: publish a live demo of the "needle playground". It's small enough that it should be pretty cheap to run this on a little VPS somewhere!

Show HN: Needle: We Distilled Gemini Tool Calling into a 26M Model

https://github.com/cactus-compute/needle
48•HenryNdubuaku•1h ago•9 comments

Show HN: Agentic interface for mainframes and COBOL

https://www.hypercubic.ai/hopper
27•sai18•2h ago•8 comments

Show HN: Gigacatalyst – Extend your SaaS with an embedded AI builder

22•namanyayg•3h ago•8 comments

Show HN: Statewright – Visual state machines that make AI agents reliable

https://github.com/statewright/statewright
22•azurewraith•5h ago•6 comments

Show HN: Cook a Django project well, the agent-skill take on cookiecutter

https://github.com/RobustaRush/seedkit
2•kmmbvnr_•1h ago•0 comments

Show HN: OpenClaw OS – OSS Claude Cowork Built on Top of OpenClaw

https://github.com/thesysdev/openclaw-os
2•zahlekhan•1h ago•0 comments

Show HN: TikTok but for scientific papers

https://andreaturchet.github.io/website/index.html
172•ciwrl•1d ago•68 comments

Show HN: A modern Music Player Daemon based on Rockbox firmware

https://github.com/tsirysndr/rockbox-zig
111•tsiry•3d ago•24 comments

Show HN: One-shot NAT traversal library

https://warpgate.io/
3•Uptrenda•2h ago•0 comments

Show HN: Profine - Profile and rewrite your PyTorch training loop on real GPUs

https://github.com/ProfineAI/profine-cli
3•aisinghal•2h ago•0 comments

Show HN: OpenGravity – A zero-install, BYOK vanilla JS clone of Antigravity

https://github.com/ab-613/opengravity
100•ab613•23h ago•30 comments

Show HN: Grunden – Frontier AI inference hosted in Sweden, OpenAI-compatible

https://grunden.ai
4•fsrc•4h ago•1 comments

Show HN: E2a – Open-source email gateway for AI agents

https://github.com/Mnexa-AI/e2a
43•mnexa•23h ago•3 comments

Show HN: Doomscroll the Goverment's UFO Files as One Gigantic Microfilm

https://hypergrid.systems/war.gov-ufo-viewer/microfilm2?page=2
15•keepamovin•15h ago•7 comments

Show HN: Formal Verification with Lean

https://www.daniellowengrub.com/blog/2026/04/30/lean
3•lowdanie•6h ago•0 comments

Show HN: Is Github Online?

https://isgithub.online/
5•vednig•6h ago•2 comments

Show HN: How Scaleway brought the first RISC-V servers to the cloud

https://www.scaleway.com/en/blog/risc-v-servers-in-the-cloud/
4•enthusaist•7h ago•0 comments

Show HN: Music visualizers that react to audio in real time

https://vizz.fm/app/
6•lowtecky•7h ago•2 comments

Show HN: I mage GhosttyFX, a JavaFX terminal view that uses libghostty

https://github.com/vlaaad/ghosttyfx/
3•vlaaad•8h ago•1 comments

Show HN: An index of indie web/blog indexes

https://theindex.fyi
145•rocketpastsix•2d ago•40 comments

Show HN: I made a Clojure-like language in Go, boots in 7ms

https://github.com/nooga/let-go
270•marcingas•3d ago•83 comments

Show HN: Countries where you can leave your MacBook at a random coffee shop

https://vouchatlas.com
71•canergl•2d ago•87 comments

Show HN: Rust but Lisp

https://github.com/ThatXliner/rust-but-lisp
207•thatxliner•2d ago•72 comments

Show HN: adamsreview – better multi-agent PR reviews for Claude Code

https://github.com/adamjgmiller/adamsreview
81•adamthegoalie•1d ago•50 comments

Show HN: Safe-install – safer NPM installs with trusted build dependencies

https://www.npmjs.com/package/@gkiely/safe-install
17•gkiely•19h ago•3 comments

Show HN: Java/Spring Boot Idempotency Library

https://github.com/josipmusa/idempotency4j
2•anaq42•11h ago•1 comments

Show HN: Building a web server in assembly to give my life (a lack of) meaning

https://github.com/imtomt/ymawky
422•imtomt•2d ago•223 comments

Show HN: I built a fair-price checker App for home repairs using BLS wage data

https://apps.apple.com/us/app/quotebuster/id6765996986
4•manojtrek•12h ago•0 comments

Show HN: TRUST – Coding Rust like it's 1989

https://github.com/wojtczyk/trust
170•wojtczyk•5d ago•87 comments

Show HN: Tessera – Turn coding agent sessions into structured work

https://github.com/horang-labs/tessera
4•horanglabs•14h ago•0 comments