The palOMine appliance
One box. One config. Plug-and-play.
palOMine ships pre-configured on a single GMKtec EVO X-2 with 128 GB of unified memory: coding agent, voice, image generation, chat gateways, retrieval, and memory — all running on open-weight models over vendor-neutral hardware. No inference API keys. No cloud inference required. No rack, no GPU mix to size, no MLOps team to staff.
Five surfaces, one bounded-authority core.
Not a toolkit, not a sandbox — every surface below is wired, the accounts are mine, and the rules are mine. You decide what it’s allowed to touch; remote requests can’t mutate anything by default.
- Coding agent
A local coding agent that reads, edits, and runs. No API keys, no vendor rate limits.
- Voice in & out
Speech-to-text and text-to-speech on the box. Listen back, dictate, hands-free.
- Image generation
Open-weight image generation — your prompts and outputs, on-device.
- Chat gateways
Bridge to Telegram, Discord, Signal, and a terminal-grade web UI. One core, every surface.
- Retrieval & memory
Local embeddings, retrieval, and a memory that decays what you don't use and flags contradictions.
GMKtec EVO X-2, 128 GB.
One model, one configuration. Open-weight models run on a single preconfigured mini-PC — no proprietary stack, no build-to-order rack, no vendor lock-in.
- Model
- GMKtec EVO X-2, 128GB unified memory — the single SKU
- Form factor
- Mini-PC desktop: sit it on a desk or shelf, plug in power + ethernet
- Compute
- Integrated GPU, NPU, and CPU are used as independent compute lanes. Nemotron handles primary reasoning and agent work on the iGPU, while Qwen3.5-2B runs on the NPU for lightweight background inference. Supporting workloads use the CPU and iGPU where appropriate.
- Quantization
- Curated per model as part of the appliance configuration; no customer sizing or quantization decisions required.
- Models
- Reasoning/coding: NVIDIA Nemotron 3.5 Lightning 30B-A3B on the iGPU. Background inference: Qwen3.5-2B on the NPU. Plus Whisper Large v3 Turbo and Moonshine Medium Streaming for speech-to-text, Kokoro for text-to-speech, FLUX.2 Klein for image generation, RealESRGAN x4+ for upscaling, nomic-embed v1 for embeddings, and bge-reranker v2 m3 for reranking. All pre-loaded as part of the appliance configuration, not a swap-in model catalogue.
- Network
- Fully offline; nothing leaves the appliance unless you explicitly forward traffic
- Monitoring
- Local logs + status page on the box itself — no third-party agent or vendor sink
- Data plane
- Prompts, embeddings, and run state stay on-device; local storage encrypted at rest
One appliance. Multiple inference lanes.
palOMine doesn't make its largest model do every job. Primary reasoning and coding run through Nemotron on the integrated GPU. Qwen3.5-2B runs independently on the NPU for lightweight background work. Speech, retrieval, image generation, and other supporting workloads use the CPU and iGPU where appropriate. The goal isn't simply to run multiple models. It's to keep expensive reasoning capacity available for the work that actually needs it.
Sustained ~57 tok/s through a long-running agent workload.
This isn't a one-prompt throughput test. In our latest agent workload, Nemotron completed 36 inference turns and generated more than 36,000 output tokens while the active working context grew beyond 40,000 tokens. Across substantive generations, the model sustained approximately 57 tok/s. A single turn generated 9,770 tokens at 57.26 tok/s. Another generated 4,682 tokens at 56.46 tok/s. More importantly, performance remains stable as an agent session grows. Generation stays around 57–58 tok/s through much of the 25K–40K context range and remains in the mid-50s beyond 40K. When context returned to approximately 11K tokens, generation returned to 59.8 tok/s.
- Workload model
- Nemotron 3.5 Lightning 30B-A3B MXFP4
- Sustained generation
- ~57 tok/s across a long-running agent workload, including multi-thousand-token generations.
- Prompt processing
- Approximately 800–1,050 tok/s on larger incremental prompt additions during the agent run.
- TTFT — cached agent workload
- ~0.59s median. Ordinary incremental turns generally return to generation in roughly 0.3–0.9 seconds.
- Long-context behavior
- The active agent context grew beyond 40K tokens while generation remained in the mid-50 tok/s range.
- Context caching
- Consecutive turns commonly reuse 97–99%+ of the existing context. palOMine processes the changed portion of the working context instead of repeatedly prefilling the entire conversation.
- Independent NPU inference
- Qwen3.5-2B sustains approximately 25 tok/s on the NPU, providing a separate lane for lightweight agent tasks without consuming the primary reasoning model’s iGPU inference capacity.
Reserve a unit for the first production run.
One email and you’re on the waitlist. We notify signups ahead of the first hardware run.