Cactus Compute押注2600万参数小模型可运行手机AI代理

做的是什么生意

Cactus Compute builds an open-source, cross-platform AI inference engine (Cactus) for phones, wearables and robots, plus Cactus Chat and the Needle model family — small tool-calling models that run fully on-device.

启动资金：Not disclosed; company is Y Combinator Summer 2025-backed per its YC company page.

起因

Co-founders Roman Shemet and Henry Ndubuaku met through YC co-founder matching in London; frustrated that nothing ran agentic AI on low-cost phones, they built Cactus, a cross-platform inference engine with Flutter, React Native and Kotlin bindings that runs Hugging Face models locally with cloud fallback. The company was founded in 2025 and went through YC Summer 2025.

经过

Needle came from a bet inside the company: tool calling is retrieval-and-assembly, not reasoning, so massive models are overkill. The team pretrained a 26M-parameter 'Simple Attention Network' with no FFN layers on 200B tokens across 16 TPU v6e in 27 hours, then post-trained it for 45 minutes on Gemini-synthesized tool-calling data. The result was a 14MB model doing 6,000 tok/s prefill and 1,200 tok/s decode on consumer devices, beating FunctionGemma-270M, Qwen-0.6B, Granite-350M and LFM2.5-350M on single-shot function calling. It launched MIT-licensed on 2026-05-12, drew 776 points and 211 comments on HN, and Gigazine noted the wrinkle that Google's Gemini API terms prohibit distillation.

结果

Still live and scaling: as of September 2026 the repo ships Needle 2, a 45M-parameter model in a single 14MB binary with 2-bit Cactus Quants, confidence-gated tool calls and an arXiv paper (2607.18363); YC lists an 8-person active SF team with Cactus Engine, Needle and Cactus Hybrid products.

可借鉴之处

将问题缩小至小模型能胜出：将工具调用与推理分离，发布击败270M对手的26M模型，并把HN 776分的发布转化为商业引擎的分发渠道。