Introducing the Developer Edition: local coding models you own
Published June 28, 2026 · 5 min read
AI coding assistants have quietly become the most expensive — and most data-hungry — tool in the stack. Every keystroke of autocomplete ships your source to someone else's servers and meters you for the privilege. The AgentBox Developer Edition flips that: open coding models, running on hardware you own, with nothing leaving the building.
Why a dedicated coding box
Two problems pushed us here. First, cost: completion endpoints fire thousands of times a day per developer, and metered billing compounds with your team's productivity. Second, exposure: most coding assistants are a third-party processor sitting between your engineers and your crown-jewel IP. For regulated teams, that's a non-starter.
Owning the metal solves both at once. The purchase is one-time, so heavy use makes the economics better, not worse. And because inference is local, your code simply never leaves your perimeter.
The model lineup
Every model is open-weight and pre-tuned for the RK3588 NPU. Swap between them from the appliance console — no re-imaging, no cloud round-trip.
CodeGemma, from Google's Gemma family, is a strong everyday default with excellent fill-in-the-middle completion. Qwen2.5-Coder leads most local accuracy comparisons for its size. DeepSeek-Coder-V2-Lite brings broad language coverage and long-context repo reasoning, and StarCoder2 offers a permissive, transparent base for fine-tuning on your own repos.
Yes, you can run Kimi — on a cluster
People keep asking for Kimi, so let's be precise. Moonshot AI's Kimi-Dev is a 72B software-engineering model, and Kimi K2 is a roughly trillion-parameter mixture-of-experts. Neither fits on a single node — and we won't pretend otherwise. Instead, AgentBox pools aggregate memory across a cluster, and the mesh coordinates the shards so a multi-node AgentBox can host Kimi-Dev. It's a clean illustration of why the 1→16 node architecture exists.
Drop it into your editor
The box exposes an OpenAI-compatible endpoint, so VS Code, JetBrains and your CI agents point at it and just work. Ollama, llama.cpp and a container runtime are pre-wired; first completion lands in about five minutes.
Launch pricing
To get the Developer Edition into developers' hands, we're selling founding-batch units at cost: $899 (full price $1,798). It ships Q4 2026, and reservations are fully refundable until your unit ships.
Developer Edition FAQ
More questions? See the full FAQ or talk to us.