https://llmobile.news/companies/cactus/A concise news ticker covering AI on mobile devices, local models, apps and hardware.en-GBhttps://llmobile.news/ticker/cactus-needle-3/https://llmobile.news/ticker/cactus-needle-3/Thu, 17 Sep 2026 00:00:00 +0200Cactus has released Needle 3, a model with 121 million parameters for tool calling, structured extraction and text embedding. The whole model ships as a single binary of 8 to 29 MB, and Cactus says it beats models ten times its size on mobile tool calls and matches models two to three times larger on extraction. The README names mobile devices, wearables, robots, smart home, automotive and microcontrollers as targets. Needle 3 uses a Laddered Simple Attention Network, trained so that every depth from 2 to 20 layers yields a deployable model. Most of its parameters sit in an n-gram memory the README calls the engram, so the 121M model does the arithmetic of a 50M one. Weights are stored at 2.125 bits per weight in Cactus’s own quantisation format. A grammar compiled from the user’s schema constrains the output, and every response carries a calibrated confidence score next to the function calls and the model’s reasoning. The 20-layer, 121M variant scores 86.0% exact-call accuracy on Mobile Actions in the chart Cactus published with the README. That is ahead of LFM2.5 1.2B at 82.4%, Qwen3.5 0.8B at 76.0% and Needle 2 at 63.5%, while DeepSeek V4 Flash reaches 88.4%. On DroidCall the same variant scores 47.0% against 35.5% for LFM2.5 1.2B, and on BFCL v4 it scores 50.2% against 62.0% for LFM2.5 1.2B. The 8-layer, 52M variant reaches 36.8% on Mobile Actions and the 4-layer, 29M variant 11.7%. Fine-tuning on the DroidCall dataset lifts every subnetwork by 18 to 36 points, according to the README, and tuned subnetworks from 4 layers upward exceed DeepSeek V4 Flash. Fine-tuning and export run locally at 4-bit with the needle finetune and needle build commands. Cactus offers advanced 2-bit post-training through its own platform. Developers install the Python package with pip install cactus-needle and declare tools as decorated functions. Each deployment target gets a prebuilt engine under 1 MB that loads the model file at startup, with builds for macOS ARM64, Linux ARM64, the browser and WASI. The weights are on Hugging Face under Apache 2.0, according to the model card, and a browser demo is on the Cactus website. Telemetry is on by default and can be switched off with the NEEDLE_TELEMETRY=0 and DO_NOT_TRACK=1 environment variables. Source: https://github.com/cactus-compute/needle Read the article: https://llmobile.news/ticker/cactus-needle-3/Tool callingQuantisationOpen source