Releases·Model Releasefeatured
Cactus Needle 3: A Sliceable 8-29MB Automation Foundation Model That Matches DeepSeek v4 Flash
Cactus Compute released Needle 3, a 121M-parameter foundation model for tool calls, structured extraction, and embeddings that runs fully on-device with no network. It scores 86.0 on Mobile Actions, approaching DeepSeek V4 Flash's 88.4 while beating LFM2.5 1.2B and Qwen3.5 0.8B. A single weight set supports every depth from 2 to 20 layers, letting deployers pick 29-121M parameter subnetworks (8-29MB binaries) for phones, watches, and microcontrollers.
248 engagement1 sourceThu, Sep 17, 2026, 08:05 PM
AI Overview
- Cactus Compute released Needle 3, a 121M-parameter on-device automation model trained on 360B tokens that scores 86.0 on Mobile Actions versus DeepSeek V4 Flash's 88.4.
- Needle 3 uses a Simple Attention Network with a Monarch Hadamard MLP and 70.8M hashed n-gram engram parameters, running at 100 MFLOPs per token versus 296 for a same-shape transformer.
- One weight set slices into deployable models at every depth from 2 to 20 layers; LoRA-fine-tuned subnetworks gain 18-36 points on DroidCall, with 29M and up passing DeepSeek V4 Flash.
- The sub-1MB CPU engine reads 2.125-bit Cactus Quants weights in place from a single 8-29MB file, decoding up to 4k tokens/s on a Raspberry Pi 5 with calibrated confidence gating.
- Needle 3 grounds every argument as a span of the request, emits calls under a byte-level schema grammar, and returns an empty list when no declared tool can serve a request.
Related
Releases·Model Releasefeatured
Cactus releases Needle2, a 14MB agentic LLM for edge devices
Cactus released Needle2, a 14MB agentic LLM for tool call, device use, and structured extraction on phones, wearables, smart homes, small robots, and microcontrollers. The model runs a full session in 28MB of RAM and achieves 500 tokens/sec on a Raspberry Pi 5.
Cactus released Needle2, a 14MB agentic LLM for tool call, device use, and structured extraction on phones, wearables, smart homes, small robots, and microcontrollers. The model runs a full session in 28MB of RAM and achieves 500 tokens/sec on a Raspberry Pi 5.
Updated Sun, Aug 16, 2026, 10:19 AM — MimiModel, a C inference engine, runs Needle2 on a $5 ESP32-S3 chip.
120135 engagement3 sourceshackernews, reddit, githubMon, Aug 10, 2026, 05:22 PM
Releases·Model Release
Cactus releases Needle, a 26M parameter function calling model
Cactus Compute released Needle, a 26 million parameter model specialized for function calling, claiming it outperforms much larger models. The model is open-source with a demo on Hugging Face and code on GitHub.
Cactus Compute released Needle, a 26 million parameter model specialized for function calling, claiming it outperforms much larger models. The model is open-source with a demo on Hugging Face and code on GitHub.
2.6k engagement1 sourceyoutubeSun, Jul 12, 2026, 01:00 PM
Social·Community
Users report LFM2.5 2.6B underperforms Qwen3.5 4B on tool calling in OpenCode
A developer tested Liquid AI's LFM2.5 2.6B dense model on agentic tool calling and coding with OpenCode, finding it vastly underperformed Qwen3.5 4B on tool calling. The model struggled with actual tool calls and was confused about the working directory, despite running at ~90 t/s with Q8 on an M5 Pro.
A developer tested Liquid AI's LFM2.5 2.6B dense model on agentic tool calling and coding with OpenCode, finding it vastly underperformed Qwen3.5 4B on tool calling. The model struggled with actual tool calls and was confused about the working directory, despite running at ~90 t/s with Q8 on an M5 Pro.
24 engagement1 sourceredditWed, Aug 5, 2026, 06:15 PM
Products·launched
Cactus Needle 2: fine-tunable 2-bit LLM for intent matching
Cactus Needle 2 is a 2-bit quantized language model that can be fine-tuned for specific tasks like intent matching, aiming to match DeepSeek v4 on such tasks. It is provided with a playground for experimentation and a Python library for fine-tuning, targeting developers who need efficient, customizable models.
Cactus Needle 2 is a 2-bit quantized language model that can be fine-tuned for specific tasks like intent matching, aiming to match DeepSeek v4 on such tasks. It is provided with a playground for experimentation and a Python library for fine-tuning, targeting developers who need efficient, customizable models.
7 engagement1 sourceredditThu, Aug 20, 2026, 08:15 PM
Products·launched
CUA-S1: small 'System One' model for local computer-use decisions
CUA-S1 is a small model from the Cua team that makes low-level computer-use decisions (filling fields, checking boxes, ignoring elements) instead of a general-purpose LLM. It complements large LLMs, which remain responsible for planning and recovery. It targets developers building computer-use agents who want cheaper, faster execution.
CUA-S1 is a small model from the Cua team that makes low-level computer-use decisions (filling fields, checking boxes, ignoring elements) instead of a general-purpose LLM. It complements large LLMs, which remain responsible for planning and recovery. It targets developers building computer-use agents who want cheaper, faster execution.
1 engagement1 sourcehackernewsSat, Sep 19, 2026, 03:52 PM