Meet Needle 2: An Open 45M-Parameter Tool-Calling Model That Ships as a 14MB Binary and Runs a Full Session in 28MB of RAM

Cactus Compute has put out Needle 2, a 45M-parameter open model built for tool calling, device use, and structured extraction. The packaging is the punchline: the whole thing arrives as a single 14MB binary, and a full session reportedly sits around 28MB of RAM. That makes this feel less like a chatbot and more like a tiny utility engine that happens to speak JSON.

The company is making a blunt argument about the job it wants this model to do. If the task is turning a messy sentence into a typed function signature, Cactus says you do not need world knowledge or long-form generation. Needle 2 is meant for hardware without a GPU or an NPU, and it ships as prebuilt binaries plus a static library for macOS, Linux, Windows, Android, iOS, watchOS, tvOS, and WebAssembly.

The runtime story is just as unusual. Weights are trained and deployed at CQ2-bit with Cactus Quants, and the model stays inside Cactus’s own C++ engine, so there is no separate runtime to install and no download at inference time. The engine checks the CPU at startup and picks a kernel tier, while a byte-level grammar compiled from JSON schemas limits what tokens can appear. For structural tokens, the system can skip up to 98% of the vocabulary projection.

Cactus also keeps the memory curve flat. The attention window is 256 tokens, the system turn and tool declarations are pinned as KV sinks, and that is why memory stays near 28MB no matter how long the conversation goes. If there are five tools or fewer, they render directly. Above that, a retrieval head scores the schemas and only the top five make it through. Anything else is unreachable, not just unlikely.

On benchmarks, Needle 2 looks specialized rather than universal. Cactus says it leads both Seal-Tools splits, with 32.6 on the in-domain test and 28.7 out of domain, and it posts 63.7 on Mobile Actions. It trails on BFCL v4 overall at 42.6, which the team ties to training data centered on consumer device actions rather than general or enterprise APIs. The model also returns a confidence value for every response, and off-topic requests get an empty call instead of a forced answer.