Frontier LLMs

Cactus Compute Unveils Needle 3: A Specialized 121M Parameter Automation Model

Cactus Compute has released Needle 3, a highly specialized foundation model designed exclusively for automation, tool calling, and structured data extraction. By trading general chat capabilities for task-specific efficiency, the 121M-parameter model achieves performance competitive with much larger models on mobile action benchmarks.

The AI Multiverse Editorial Desk•Thursday, September 17, 2026•3 min read

Reported viaReddit LocalLLaMA (Model Leaks & Local AI)

Reader:

Executive Brief & 3-Point Takeaways

Needle 3 is a 121M-parameter model specialized for automation and tool calling, not conversational chat.

The architecture uses a Monarch Hadamard MLP and hashed n-gram tables to reduce computational overhead.

The model achieves 86.0 on Mobile Actions benchmarks, outperforming several larger models at f16 precision.

Impact:

Needle 3 demonstrates that specialized, small-scale architectures can outperform massive general-purpose models in specific automation tasks, enabling high-performance, on-device AI without cloud dependency.

Fact-checked against primary sourcesCategory: Frontier LLMs

Cactus Compute Unveils Needle 3: A Specialized 121M Parameter Automation Model

Illustration: AI-generated

Key Findings & Takeaways

Needle 3 is a 121M-parameter model specialized for automation and tool calling, not conversational chat.
The architecture uses a Monarch Hadamard MLP and hashed n-gram tables to reduce computational overhead.
The model achieves 86.0 on Mobile Actions benchmarks, outperforming several larger models at f16 precision.

Why It Matters

Key Details

Cactus Compute has introduced Needle 3, a 121M-parameter model optimized for function calling and structured data extraction. Unlike general-purpose LLMs, Needle 3 is designed to operate entirely on-device, eliminating the need for network connectivity. It is available via Hugging Face, GitHub, and PyPI.

Technical Architecture

Needle 3 utilizes a 'Simple Attention Network' architecture, replacing traditional dense feed-forward layers with a Monarch Hadamard MLP. This design significantly reduces arithmetic requirements—requiring only 100 MFLOPs per token compared to 296 MFLOPs for standard transformers of similar shape. A substantial portion of the model's knowledge is stored in hashed n-gram tables (engrams), which are accessed via gather operations rather than computation.

What to Expect

The model is strictly focused on automation. It processes requests by mapping them to function calls with validated arguments, ensuring JSON output that adheres to predefined schemas. In benchmark testing, the 20-layer version of Needle 3 scored 86.0 on Mobile Actions, placing it within reach of DeepSeek V4 Flash (88.4) while maintaining a significantly smaller footprint.