Enki Multimedia

Engineering lab

Open-source AI Systems Lab.

Where we build and test open-source software for AI systems on our own hardware: a local coding agent on distributed inference, durable memory for agents, and the infrastructure that connects them. We publish what we learn as we go.

The lab builds on more than twenty years of open-source work in Python, Erlang, networking, runtimes and distributed systems.

Setup verified on 16 November 2026.

What we are working on

A local coding agent

One capable coding agent, served from our own machines, with the model split across two nodes through distributed inference.

Barrel Agent Memory

Durable memory, hybrid retrieval and context management for agents, developed and tested here against a real agent workload.

Barrel is developed by Enki Bridge, a separate company that uses Enki Multimedia's engineering services.

Distributed AI infrastructure

Experiments with networking, scheduling and serving across machines, and with hardware from different manufacturers.

What runs today

The setup as verified on 16 November 2026. It will change as the experiments move on.

HardwareTwo ASUS Ascent GX10 systems, each with an NVIDIA GB10, 128 GB-class unified memory, 20 ARM CPU cores and 1 TB of NVMe storage.
InterconnectTwo active 200 Gb/s ConnectX-7 interfaces per machine. NCCL communication over RoCE is confirmed on both interfaces.
InferenceGLM-5.3-Flash, served across both machines with TensorFold.
AccessAn OpenAI-compatible inference API, for our agents and tests.
Supporting host (planned)The model uses most of the memory on both machines, so agent orchestration, Barrel Agent Memory and test services will run on a separate host.
Supporting host (planned)Agent orchestration, Barrel Agent Memory, tests
GX10 node 1NVIDIA GB10, serves part of the model
GX10 node 2NVIDIA GB10, serves part of the model
Roles, not a network map: we do not publish addresses or internal configuration.

What we have not measured yet

This page lists only what we have checked. The results below are not available yet; when we publish them, they will come with the setup and scripts needed to reproduce them.

  • Coding benchmarks for the local agent
  • Sustained throughput across the two interconnect links
  • Long-context inference under concurrent load

Lab notes

Experiments and findings, newest first. Each entry will link to its write-up and code.

16 November 2026Two-node setup verified: dual-node inference with TensorFold, NCCL over RoCE on both links, and the OpenAI-compatible API in service.

Built on our open-source work

The lab continues work we have published for more than twenty years: serving Python applications, Erlang HTTP clients and servers, QUIC and other network protocols, embedded runtimes, and distributed systems and databases. The projects here are the ones the lab builds on most directly; the full catalogue is on our open-source page.

Hardware partnerships

We would like to work with AMD, NVIDIA, ASUS and other manufacturers on:

  • Evaluation hardware for the experiments on this page
  • Independent testing, published with its limits
  • Cross-platform development, so the same software runs on different accelerators
  • Reproducible technical findings: setups, scripts and results others can check

Results are published on this page whoever supplied the hardware.