5acAI-Lab. An open-source AI lab you can actually own.
The future of business AI shouldn't sit on foreign APIs and closed licenses. We build our own stack — open, runs in place, self-owned. Fork it. Use it. Improve it. Own it.
Why 5ac runs an AI lab
5ac doesn't just run OPC (G-Company OS) for businesses — it's also an open lab where serving, tuning and benchmark recipes are rewritten, measured and published openly. Every recipe in the repo runs for real on hardware we own — not on slides.
Open, runs in place, self-owned — that's how we believe AI should reach businesses.
The two pillars of 5acAI-Lab
OPC — G-Company OS
One commander, a ready-to-work AI team. The outcome-first layer — it tells your business the results, not the plumbing.
outcome-first · homepage layerLocal AI Engine
An on-prem inference cluster inside your server room — data never leaves the building. The infrastructure layer, built for technical buyers.
on-prem · infrastructureTwo layers, two languages. OPC buyers hear outcomes; builders following the Lab hear the engineering. We don't mix them.
Repository & Recipe
All public recipes live on GitHub under an MIT license. We fork, audit and localize from the global open-source ecosystem — then re-measure on our own hardware.
- Production serving recipe (tensor-parallel 2, speculative decoding, long context) — full ops scripts plus a fail-closed hotfix.
- On-device tuning recipe — serving with speculative decoding, tuned in place on the box.
- Ops scripts: start / stop / status / logs / smoke-test, reboot-safe systemd units.
- Live verifier: Responses-API 4-gate checks (text/SSE, tool continuation, JSON schema, multi-turn prefix reuse).
Benchmark — measured, not slides
Every figure below was measured on real on-prem hardware, n=5, on the same weight set.
| Workload | DSpark (block-7) | MTP (EAGLE 3/1/4) | Winner |
|---|---|---|---|
| Code — LRUCache + test | 51.5 tok/s | 34.5 tok/s | DSpark ~1.5× |
| Short chat (thinking on) | 23.2 tok/s | 21.0 tok/s | DSpark, slightly |
| Long essay (Babbage → GPUs) | 18.3 tok/s | 24.1 tok/s | MTP ~1.3× |
Lab takeaway: use DSpark for agents, code and everyday chat; MTP only when writing long essays. Tuning measured: GDN state bf16 (+3% vs float32), spec steps 3/1/4 hit the peak, pin 10 performance cores (+2–7% decode).
(Source: MiaAI-Lab + Kubesimplify benchmark, 14–18 Aug 2026 — confirmed by the CTO before publishing.)
Roadmap
- P0 (in progress): Lab landing page, public recipe repo, two technical blog posts (production serving · speculative decoding vs MTP).
- P1: on-device tuning, Local AI Engine FAQ + pricing, Vietnamese operations documentation.
- P2: real on-prem handover case study, "build it yourself vs hire an agency".
Does your business need on-prem?
We install and hand over an AI cluster inside your server room — data never leaves the building, and you own the entire stack. Packages from VND 175M, with an operations care plan.
Help build the Lab
Every recipe needs testers, contributors and critics. Join the 5ac community to measure benchmarks together, review hotfixes and write recipes.