{"schema_version":"1.0","service":"Publicasta","type":"article","id":327,"slug":"needle2_14mb_action_model_robots_2026_08_15","title":"Needle 2 puts a 14 MB action model near robots — and raises the safety bar","excerpt":"Cactus Compute’s tiny tool-calling model points to faster private control for devices, but physical robots need refusals, policies and confirmations as much as speed.","language":"en","default_language":"en","canonical_url":"https://publicasta.com/robots/needle2_14mb_action_model_robots_2026_08_15?lang=en","image":{"url":"https://publicasta.com/storage/projects/11/pages/327/2026/08/c9600b61-707f-4ae3-aa45-25de76d984ba.webp","alt":"On-device AI hub coordinating a home robot, phone, wearable and smart-home tools without cloud access"},"publisher":{"id":11,"slug":"robots","name":"Robots and Autonomous Systems","url":"https://publicasta.com/robots"},"author":{"name":"Anton R"},"published_at":"2026-08-15T07:49:36+00:00","updated_at":"2026-08-15T07:49:36+00:00","content_markdown":"Cactus Compute’s Needle 2 release is easy to file under “another tiny language model”. That would miss the robotics story. The company describes Needle 2 as an open 45-million-parameter model for tool calling, device use and structured extraction, packed into a single 14 MB binary with about 28 MB of session RAM. The launch page points directly at phones, wearables, connected home devices, Raspberry Pi boards, newer microcontrollers and small robots. In other words, it is aimed at machines that may need to turn a spoken or typed request into a physical action without waiting for a cloud model.\n\n ![On-device AI hub coordinating a home robot, phone, wearable and smart-home tools without cloud access](https://publicasta.com/storage/projects/11/pages/327/2026/08/c9600b61-707f-4ae3-aa45-25de76d984ba.webp) The interesting question is not whether a 45M model can chat like a frontier assistant. It cannot, and it does not need to. A useful robot often needs a narrower skill: map “turn the hallway lights down”, “open the gripper halfway”, “log this sensor reading”, or “ask before unlocking the door” into a safe function call with the right parameters. Needle 2 is part of a wider shift from general chatbots toward small local action models. That shift can make robots faster, cheaper and more private, but it also moves model mistakes from text into the physical world.\n\n ## What Cactus says it built\n\n The official Cactus page presents Needle 2 as a model for tool calling, device use and structured extraction. It says the model is built on a Simple Attention Network, compressed with Cactus Quants to CQ2-bit form and shipped as its own engine. The headline numbers are deliberately tiny by current AI standards: 45 million parameters, a 14 MB binary, roughly 28 MB peak session RAM, and claimed decode speeds of about 500 tokens per second on a Raspberry Pi 5, 400 to 1,500 tokens per second on VR devices such as Meta Quest 3S and Apple Vision Pro, and 300 to 700 tokens per second on some sub-$200 phones.\n\n Those numbers should be read as vendor claims until independent tests reproduce them across devices, schemas and real workloads. Still, the packaging target matters. The GitHub repository describes Needle as a 14 MB foundation model for tiny devices including phones, wearables, smart home and robots. The Cactus runtime repository describes quantization, kernels and inference for mobiles, wearables, smart home and robots. Hugging Face lists Cactus-Compute/needle2 with tags for tool calling, function calling, on-device, edge, quantization and WebAssembly, and shows the model as a text-generation artifact tied to arXiv:2607.18363. This is not merely a slide about future hardware; the code, model page and runtime positioning all point at edge deployment.\n\n ## Why robots need this layer\n\n Most robots do not need an encyclopedic conversation partner in the control path. A domestic robot, inspection rover, wearable assistant or smart-home hub needs a compact interpreter between people, sensors and actuators. Classical intent parsers and rule trees can do this when commands are predictable. Large cloud models can do it when latency, privacy, connectivity and inference cost are acceptable. The gap is the everyday device that must work quickly, locally and cheaply, with a command space broader than a fixed menu but narrower than open-ended reasoning about the world.\n\n That gap is exactly where small tool-calling models become interesting. They can translate natural language into structured calls, extract parameters from messy phrasing and decline irrelevant requests if they are trained and constrained well enough. A robot arm does not need to write an essay before moving to a safe pose; it needs to select the right allowed action, fill the distance or object name, and know when the request is ambiguous. A home controller does not need to debate philosophy before changing a thermostat; it needs to know whether the user asked for temperature, lighting, security, music or nothing actionable.\n\n ## The Hacker News reaction found the nerve\n\n The Show HN thread for Needle2 drew strong developer interest, with hundreds of points and a long technical discussion. That attention is useful as a signal, not as proof of performance. The thread also surfaced the hard part: users posted examples where the web demo appeared to pick questionable or nonsensical tools, including discussion around a `lock_door` action on irrelevant input and confusion over phrases such as making a room “as dark as possible”. Individual demo failures do not prove the model is unusable. A web playground may expose a toy tool schema, lack production guardrails or run a configuration different from an embedded product.\n\n But the examples are still the right warning. In a chatbot, a bad tool choice may create an embarrassing answer or a failed API request. In a robot, a bad tool choice can move a gripper, heat a room, unlock a door, start a motor or silence an alarm. Tool calling is not just output formatting when the tool controls the environment. The model’s ability to return JSON is the beginning of the safety problem, not the end.\n\n ## Structured output helps, but semantics remain hard\n\n Developers often protect tool calling with schemas, grammars and constrained decoding. Those controls are valuable. They can prevent malformed JSON, restrict parameter types, require enumerated tool names and make the model’s output easier to validate. They cannot by themselves decide whether the user’s request should trigger any action at all. If a command is off-topic, malicious, ambiguous or missing context, the safest answer may be an empty action, a clarification question, or an escalation to a larger model or human confirmation.\n\n This distinction is crucial for robotics. A grammar can force `tool_name` to be one of five options; it cannot guarantee that “it is too bright in here” should lower lights rather than close blinds, move a robot, change a display or do nothing. A parameter validator can reject a thermostat setting outside a safe range; it cannot know whether a child joking near a smart speaker has authority to unlock a door. A local action model therefore needs a policy layer around it: tool allowlists, risk classes, confidence thresholds, user identity, device state, rate limits and explicit confirmation for high-impact actions.\n\n ## Locality is a real advantage\n\n The case for a 14 MB on-device model is strong in many robotics settings. Voice commands and sensor descriptions can stay on the device instead of travelling to a remote service. The system can keep working when the network is down. Latency can fall from a cloud round trip to local inference. A maker can build a low-cost robot without paying per-request fees. A wearable or home assistant can handle routine commands without sending household context to a server. For educational robots, assistive devices and low-cost inspection machines, those economics can decide whether a product exists at all.\n\n Local inference also changes reliability engineering. A robot can keep a small command interpreter next to the controller, while using larger models only for planning, help text or ambiguous cases. If the cloud is unreachable, the device can still handle a safe subset of actions. If connectivity returns, logs and difficult cases can be reviewed. This hybrid pattern is more plausible than the fantasy that every robot either runs a giant local model or depends entirely on the cloud.\n\n ## Locality does not remove safety work\n\n The same locality can make mistakes more immediate. A cloud service often sits behind account controls, monitoring, central updates and server-side policy. A tiny local model may be embedded in many devices with different firmware versions, tool schemas and owners. If the model confidently maps noise or unrelated speech to an action, the device may act before a remote service ever sees the problem. If a vendor ships weak default schemas, thousands of products may inherit the same unsafe behavior.\n\n Robotic safety therefore cannot be delegated to the model size or the model card. It has to be designed in the product. Dangerous tools should require confirmation. Security actions should be separated from convenience actions. The model should be allowed to say “no action” and rewarded for doing so. Logs should record the input, selected tool, parameters, confidence signal and policy decision. Tests should include nonsense inputs, adversarial phrasing, children’s speech, multilingual commands, background media and sensor contradictions. A model that calls a tool correctly on neat examples is not yet ready to control a door, heater or moving arm.\n\n ## Where Needle 2 fits among alternatives\n\n Needle 2 is not the only route to local command understanding. Classical intent systems remain attractive when the command space is small and safety is paramount. Larger local models on phones or PCs can reason more broadly but require more memory, power and careful deployment. Vendor frameworks such as Apple’s on-device foundation-model work point toward richer local AI on premium hardware. Other compact models, including function-calling and edge-oriented releases from larger ecosystems, target the same pressure point from above: more parameters, more capability, more cost.\n\n Needle 2’s editorial importance is that it pushes from below. It asks how much action routing can fit into a tiny binary. If the answer is “enough for safe routine commands under strict policy”, that is a significant layer for robots. If the answer is “fast but too eager to call tools”, the release is still valuable because it exposes the benchmark the field needs: not only tokens per second and model size, but false-action rate, correct refusal, parameter accuracy, recovery from ambiguity and behavior under constrained tool sets.\n\n ## What should be tested next\n\n The next useful evidence would be practical tests on the exact devices Cactus targets. A Raspberry Pi 5 benchmark is helpful, but robot builders also need Android phones, cheap smart displays, ESP32-class boards where feasible, Home Assistant bridges, ROS prototypes and wearable-style workloads. The tests should measure end-to-end latency, memory peaks, energy use and thermal behavior, not just decode speed. They should use realistic tool schemas with safe and unsafe actions, not only toy weather or calculator functions.\n\n Safety tests should be public and repeatable. Give the model irrelevant inputs, ambiguous commands, negations, jokes, conflicting context and malicious phrases. Check whether it returns no action when it should. Check whether it asks for confirmation before security or motion tools. Check whether adding more tools makes selection worse. Check whether multilingual commands degrade gracefully. For physical devices, the most important metric may be the number of times the system refuses to act correctly, because silence is often safer than a confident wrong movement.\n\n ## The robotics takeaway\n\n Needle 2 should not be treated as a finished robot brain. It is better understood as a candidate for a small local reflex layer: a model close to the sensors and actuators that can translate ordinary commands into structured, limited, auditable actions. That layer is exactly what many everyday robots and smart devices lack. It could make a home assistant feel instant, let a wearable work privately, or give a small robot a practical command interface without a GPU.\n\n The price of that convenience is discipline. Tool calling for robots needs policies, simulation, logs, confirmations and refusal behavior. A 14 MB model that can choose a function is impressive. A 14 MB model that knows when not to choose one would be more important. The near-term breakthrough in domestic and small-field robotics may not be a humanoid body or a giant cloud brain. It may be a tiny local model surrounded by enough engineering restraint that the machine presses the right button — and leaves the dangerous ones alone.\n\n ## Sources\n\n Cactus Compute’s Needle 2 launch page and Cactus runtime material; GitHub repositories `cactus-compute/needle` and `cactus-compute/cactus`; Hugging Face model page `Cactus-Compute/needle2`; arXiv paper 2607.18363 on Simple Attention Networks; the Show HN discussion used only as a developer-reaction and failure-mode signal.","available_translations":[{"language":"ar","title":"Needle 2 يقرّب نموذج فعل بحجم 14 ميغابايت من الروبوتات ويرفع شرط السلامة","html_url":"https://publicasta.com/robots/needle2_14mb_action_model_robots_2026_08_15?lang=ar","markdown_url":"https://publicasta.com/robots/needle2_14mb_action_model_robots_2026_08_15.md?lang=ar","json_url":"https://publicasta.com/robots/needle2_14mb_action_model_robots_2026_08_15.json?lang=ar","api_url":"https://publicasta.com/api/public/v1/channels/robots/articles/needle2_14mb_action_model_robots_2026_08_15?lang=ar"},{"language":"de","title":"Needle 2 bringt ein 14-MB-Aktionsmodell zu Robotern und erhöht die Sicherheitslatte","html_url":"https://publicasta.com/robots/needle2_14mb_action_model_robots_2026_08_15?lang=de","markdown_url":"https://publicasta.com/robots/needle2_14mb_action_model_robots_2026_08_15.md?lang=de","json_url":"https://publicasta.com/robots/needle2_14mb_action_model_robots_2026_08_15.json?lang=de","api_url":"https://publicasta.com/api/public/v1/channels/robots/articles/needle2_14mb_action_model_robots_2026_08_15?lang=de"},{"language":"en","title":"Needle 2 puts a 14 MB action model near robots — and raises the safety bar","html_url":"https://publicasta.com/robots/needle2_14mb_action_model_robots_2026_08_15?lang=en","markdown_url":"https://publicasta.com/robots/needle2_14mb_action_model_robots_2026_08_15.md?lang=en","json_url":"https://publicasta.com/robots/needle2_14mb_action_model_robots_2026_08_15.json?lang=en","api_url":"https://publicasta.com/api/public/v1/channels/robots/articles/needle2_14mb_action_model_robots_2026_08_15?lang=en"},{"language":"es","title":"Needle 2 acerca un modelo de acción de 14 MB a los robots y sube la exigencia de seguridad","html_url":"https://publicasta.com/robots/needle2_14mb_action_model_robots_2026_08_15?lang=es","markdown_url":"https://publicasta.com/robots/needle2_14mb_action_model_robots_2026_08_15.md?lang=es","json_url":"https://publicasta.com/robots/needle2_14mb_action_model_robots_2026_08_15.json?lang=es","api_url":"https://publicasta.com/api/public/v1/channels/robots/articles/needle2_14mb_action_model_robots_2026_08_15?lang=es"},{"language":"fr","title":"Needle 2 rapproche un modèle d’action de 14 Mo des robots et relève l’exigence de sécurité","html_url":"https://publicasta.com/robots/needle2_14mb_action_model_robots_2026_08_15?lang=fr","markdown_url":"https://publicasta.com/robots/needle2_14mb_action_model_robots_2026_08_15.md?lang=fr","json_url":"https://publicasta.com/robots/needle2_14mb_action_model_robots_2026_08_15.json?lang=fr","api_url":"https://publicasta.com/api/public/v1/channels/robots/articles/needle2_14mb_action_model_robots_2026_08_15?lang=fr"},{"language":"pl","title":"Needle 2 zbliża 14 MB model działania do robotów i podnosi próg bezpieczeństwa","html_url":"https://publicasta.com/robots/needle2_14mb_action_model_robots_2026_08_15?lang=pl","markdown_url":"https://publicasta.com/robots/needle2_14mb_action_model_robots_2026_08_15.md?lang=pl","json_url":"https://publicasta.com/robots/needle2_14mb_action_model_robots_2026_08_15.json?lang=pl","api_url":"https://publicasta.com/api/public/v1/channels/robots/articles/needle2_14mb_action_model_robots_2026_08_15?lang=pl"},{"language":"ru","title":"Needle 2 приближает 14 МБ action model к роботам — и повышает планку безопасности","html_url":"https://publicasta.com/robots/needle2_14mb_action_model_robots_2026_08_15?lang=ru","markdown_url":"https://publicasta.com/robots/needle2_14mb_action_model_robots_2026_08_15.md?lang=ru","json_url":"https://publicasta.com/robots/needle2_14mb_action_model_robots_2026_08_15.json?lang=ru","api_url":"https://publicasta.com/api/public/v1/channels/robots/articles/needle2_14mb_action_model_robots_2026_08_15?lang=ru"},{"language":"zh","title":"Needle 2 把 14 MB 行动模型带到机器人旁边，也抬高了安全门槛","html_url":"https://publicasta.com/robots/needle2_14mb_action_model_robots_2026_08_15?lang=zh","markdown_url":"https://publicasta.com/robots/needle2_14mb_action_model_robots_2026_08_15.md?lang=zh","json_url":"https://publicasta.com/robots/needle2_14mb_action_model_robots_2026_08_15.json?lang=zh","api_url":"https://publicasta.com/api/public/v1/channels/robots/articles/needle2_14mb_action_model_robots_2026_08_15?lang=zh"}],"_links":{"self":"https://publicasta.com/robots/needle2_14mb_action_model_robots_2026_08_15.json?lang=en","api":"https://publicasta.com/api/public/v1/channels/robots/articles/needle2_14mb_action_model_robots_2026_08_15?lang=en","html":"https://publicasta.com/robots/needle2_14mb_action_model_robots_2026_08_15?lang=en","canonical":"https://publicasta.com/robots/needle2_14mb_action_model_robots_2026_08_15?lang=en","markdown":"https://publicasta.com/robots/needle2_14mb_action_model_robots_2026_08_15.md?lang=en","json":"https://publicasta.com/robots/needle2_14mb_action_model_robots_2026_08_15.json?lang=en","channel":"https://publicasta.com/api/public/v1/channels/robots","channel_articles":"https://publicasta.com/api/public/v1/channels/robots/articles","search":"https://publicasta.com/api/public/v1/search","documentation":"https://publicasta.com/api-docs#reading-publicasta","openapi":"https://publicasta.com/api-docs/openapi.json","llms":"https://publicasta.com/llms.txt"}}