Cactus

Cactus Compute builds automation foundation models small enough to ship inside the device rather than call out to a data centre. Its Needle family, whose third generation landed in September 2026, compresses to binaries of 8 to 29 MB and targets phones, wearables, AR glasses, smart-home hubs, robots, cars, Macs and PCs, game consoles, TVs and microcontrollers.

Needle 3 is not a shrunken chatbot. It is trained for three jobs an on-device assistant actually needs: tool calling (picking the right function and pulling its arguments out of a user's request), structured extraction (turning loose text — an invoice, a booking confirmation, a form — into typed JSON), and text embeddings for local semantic search and matching. Cactus states it was trained on 360 billion tokens of a proprietary structured dataset oriented towards mobile automation, and claims it beats models ten times its size on mobile tool calls and matches models two to three times larger on extraction.

The architecture is the interesting part. Cactus calls it intelligence laddering: one set of weights, 29-121M Laddered Simple Attention Networks, from which subnetworks of 2 to 20 layers can each be used as an independent model. A developer picks a depth to fit the silicon they have, and the 8 MB figure is the shallow end of the same ladder that reaches 29 MB. Everything ships in CQ2, Cactus's own 2-bit quantisation format. Reported throughput on a Raspberry Pi 5 is 400-4,000 tokens/second decode and 1,000-10,000 tokens/second prefill.

Deployment coverage is unusually wide for a model vendor: macOS (ARM64), Linux (x86_64, ARM64, ARMv7, RISC-V64, MIPS), Windows (x86_64 and ARM64), Android, iOS, tvOS, watchOS, WebAssembly and WASI components. Day-to-day work happens through a Python package (needle) with decorators for function definitions and Pydantic models for extraction; native builds link against C headers and a compiled libneedle.a, a needle.exe, or the Wasm build.

Fine-tuning is where the money is. The open-source CLI trains LoRA adapters over a frozen base locally, exports 4-bit, and reports validation loss, for $0 and no account. The hosted Cactus Platform adds GPU compute, full-depth training, 2-bit export, confidence-head training and data augmentation: a $19 one-off Starter gives 30 days with 3 fine-tuning runs, 3,000 generated examples and 10 GB of storage; Pro is $99/month for 10 runs and 10,000 examples; larger allowances are quoted. Cactus says a 4-layer variant fine-tuned on the platform passes DeepSeek V4 Flash on its target tasks — a claim worth reading as task-specific rather than general. Models and datasets are published on Hugging Face, the code on GitHub, and the company is Y Combinator-backed.