AgentBox FAQ

3B-class models run on every SKU. The Pro (16–32GB) targets 7B quantized models plus embeddings, rerankers and RAG. The Developer Edition adds coding models, and AgentBox Home runs the multilingual Salamandra model. Larger models like Kimi-Dev 72B run across a cluster.

A box tuned to run open coding models locally — CodeGemma, Qwen2.5-Coder, DeepSeek-Coder-V2-Lite and StarCoder2 on one node, and Kimi-Dev 72B across a cluster. It exposes an OpenAI-compatible endpoint for VS Code, JetBrains and CI agents, ships Q4 2026, and is $899 at launch (full price $1,798).

A private household AI chatbot that runs the ethically trained, multilingual Salamandra model locally, with optional web search. It's a device you own — conversations stay on the box, encrypted. It's $349 at launch (full price $698).

Salamandra is an open, multilingual model from the Barcelona Supercomputing Center, trained on the MareNostrum 5 supercomputer using open-access and public-domain data and released under Apache 2.0. It covers 35 European languages.

Yes, on a cluster. Kimi-Dev is 72B and too large for a single node, so AgentBox pools aggregate memory across multiple nodes to host it.

For steady, repeated inference it usually is. A metered API bills every token forever; AgentBox is a one-time purchase, so the more you run it the more it saves. Try the ROI calculator with your own volume.

Yes. The struck-through figure is the full price; to launch the new editions we're selling founding-batch units at or near cost, so the launch price is a genuine discount rather than an inflated reference price.

No — $0 is due today. A reservation holds your place and founding-batch pricing, and is fully refundable until your unit ships. We email payment details before the production run.

Yes. Tell us your configuration and unit count via the contact page and we'll put together pricing, including larger 8–16 node clusters.

Data is encrypted on the device, secured by your own user account, password and PIN. AgentBox Home also offers an optional hardware Privacy Key that ties decryption to a physical key only you hold.

$99 during launch (sold at roughly cost), returning to $199 afterward. It's optional — standard encryption with user/password/PIN is included on every unit. The key seals the box's data when removed, for protection beyond a PIN alone.

Yes. AgentBox runs fully offline and can be air-gapped, so data never leaves your perimeter — a fit for HIPAA-aligned healthcare, regulated finance and locked-down industrial sites.

It boots into appliance mode with a first-run wizard. Ollama and an OpenAI-compatible API are pre-wired, so you point your existing app at the box and get a first token in about five minutes.

Yes. Any GGUF or Ollama-compatible model runs, and you can fine-tune an open base on your own data.

Nodes link over a low-latency mesh and present as a single OpenAI-compatible endpoint. Scale from 1 to 16 boxes for more parallel agents, higher throughput, or larger pooled models — with no changes to your integration code.