Quack: The DuckDB Client-Server Protocol
55 by aduffy | 3 comments on Hacker News.
Tuesday, May 12, 2026
New top story on Hacker News: Show HN: Needle: We Distilled Gemini Tool Calling into a 26M Model
Show HN: Needle: We Distilled Gemini Tool Calling into a 26M Model
10 by HenryNdubuaku | 1 comments on Hacker News.
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://ift.tt/cLsNU6K The full writeup on the architecture is here: https://ift.tt/J81foIv... 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://ift.tt/Lsay4TY ), an inference engine built from scratch for mobile, wearables and custom hardware. We wrote about Cactus here previously: https://ift.tt/M2upWRs Everything is MIT licensed. Weights: https://ift.tt/i5W3pz7 GitHub: https://ift.tt/cLsNU6K
10 by HenryNdubuaku | 1 comments on Hacker News.
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://ift.tt/cLsNU6K The full writeup on the architecture is here: https://ift.tt/J81foIv... 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://ift.tt/Lsay4TY ), an inference engine built from scratch for mobile, wearables and custom hardware. We wrote about Cactus here previously: https://ift.tt/M2upWRs Everything is MIT licensed. Weights: https://ift.tt/i5W3pz7 GitHub: https://ift.tt/cLsNU6K
Monday, May 11, 2026
Sunday, May 10, 2026
New top story on Hacker News: Ask HN: What Are You Working On? (May 2026)
Ask HN: What Are You Working On? (May 2026)
15 by david927 | 34 comments on Hacker News.
What are you working on? Any new ideas that you're thinking about?
15 by david927 | 34 comments on Hacker News.
What are you working on? Any new ideas that you're thinking about?
Saturday, May 9, 2026
Subscribe to:
Posts (Atom)