RegesARC · LIKE A KING INC.

Planning where every layer of a local LLM lives: RegesARC's memory-first approach

A Windows development preview for loading local GGUF models with explicit CPU/GPU placement, for people who run models on their own PC.

You download a large model, press load, and the app either crashes, silently falls back to a tiny context window, or runs slowly for reasons it does not explain. On a 12 GB GPU, the difference between a model that fits and one that does not is a handful of choices: how many layers go to the GPU, how big the context is, where the experts of a mixture-of-experts model live. Most tools make those choices for you. RegesARC is an attempt to show them to you first.

What is RegesARC?

RegesARC is, in the site's own words, "your local AI workstation." Today it is a Windows x64 development preview, version 0.0.37 on the download page. Its native core discovers local GGUF files, reads their metadata, plans explicit CPU and GPU placement, loads compatible models, and streams saved conversations with cancellation. It is a product of LIKE A KING INC., a small New York software studio.

The homepage also describes a larger vision: agent workflows, MCP tools and creative work. We will keep those two things apart in this article, the way the docs do.

What problem does it solve?

Local models are easy to download and hard to fit. People end up guessing at layer counts and context sizes. RegesARC aims to make memory a first-class concept: the docs describe "exact context windows," "RAM, VRAM and fit estimates," and a "Load Lab" for loading with intention. The tagline on the site, "Every byte. In its place.", is the idea in four words.

How does it work?

According to the documentation:

• Discovery. It finds GGUF files you already have, where they are. Models are not bundled.
• Planning. It reads each file's metadata and produces an explicit placement across GPU, system memory and storage, with honest memory estimates.
• Loading. It loads selected compatible models through a private Lemonade runtime, or through an exact-profile adapter called Strata for hybrid expert execution.
• Defaults. Ordinary layer offload is the default. An explicit CPU-expert option is available.
• Conversation. It streams saved conversations, supports Stop, and keeps history across relaunches.
• Boundaries. The docs describe a native host, a sandboxed renderer and an owned local runtime, with separate pages on security and privacy boundaries, offline use and validation evidence.

The documented first target is a 35B-parameter RegesCore model. The docs report that the tested artifact completed generation, cancellation, recovery, unload and saved-history relaunch in the installed app, and that the CPU-expert option and the Strata adapter each completed installed functional qualification on the exact 35B reference profile. The docs add a sentence we think is worth repeating: that qualification "does not establish faster inference than LM Studio."

The site also separates benchmarking from marketing. A docs page, "Benchmark without fooling yourself," separates prompt processing, first-token latency, decode speed and quality, and the download page says measured speeds come from one reference machine and are not promises for other hardware.

Who is it for?

People with an NVIDIA GPU and local GGUF models who want to understand why a model fits or not. The requirements: Windows 10 or 11, 64-bit. An NVIDIA GPU is recommended. The bundled GPU engine is built for GeForce RTX 40-series (CUDA sm_89) and was tested on an RTX 4070 SUPER with 12 GB. Plan on about 2.2 GB of disk space for the app plus whatever your models need.

How much does it cost?

The download is free to fetch and no price is stated. The installer is 1.28 GB (1,276,055,706 bytes), installs for the current user without administrator rights, and the page publishes a SHA-256 for version 0.0.37 so you can verify the file. Release notes and checksums are on a public GitHub releases repository.

How is it different from the obvious alternatives?

LM Studio and Ollama are mature, widely used and easy to start with. They are the right default for most people, and our own docs explicitly decline to claim we are faster than LM Studio. RegesARC's difference is the emphasis on showing and controlling placement and exact context up front, and an expert-residency approach for large mixture-of-experts models.

llama.cpp directly gives you every flag and no guardrails. RegesARC wraps a planned, inspectable placement around a runtime.

What it can't do yet

• It is a development preview. The homepage says "In development."
• The installer is not code-signed. Windows SmartScreen may warn, which is why we publish the checksum.
• Hardware coverage is narrow. The GPU engine is built for RTX 40-series, tested on one card.
• The agent platform is a plan. There is no published Python package, TypeScript SDK, public server or remote-link service in this release. The docs say a planned capability is never an installed feature.
• The 32-view interface atlas is a design simulation, not a feature inventory. The homepage says "No integration is connected in the preview" and lists "73 MCP entries" and "130 capability references & plans" as documented sources and planned adapters.
• The source code is private. The releases repository says so.
• Version labeling is inconsistent: the site download shows 0.0.37 while the docs baseline reads 0.0.1.

What's next?

The docs carry specifications for a public API, Python and TypeScript SDKs, a command-line tool, remote machines, an agentic workstation with a mode called RegesMode, and MCP and skills, all labeled as planned. The app keeps itself up to date from version 0.0.20 onward.

The business behind it

RegesARC is part of the Reges family at LIKE A KING INC. The bet is that more people will run capable models on their own hardware for privacy and cost reasons, and that fit-and-placement is the point where they get stuck. The stated plan is an inference core first and an agent workstation on top. We have no revenue, user or growth figures to report, and no price has been published. The honest differentiator today is evidence-led documentation: qualification claims are tied to what was tested, and planned features are labeled as plans.

FAQ

Is RegesARC open source?

No. The releases repository states that the source is private.

What do I need to run it?

Windows 10 or 11 (64-bit), and ideally an NVIDIA RTX 40-series GPU. It reads GGUF files you already have.

Does it come with models?

No. Models are not included.

Is it faster than LM Studio?

We do not claim that. The docs say functional qualification does not establish it.

Can I verify the installer?

Yes. The download page lists the file name, size and SHA-256.

Try it: https://regesarc.com (docs at https://docs.regesarc.com)

Written with AI assistance and reviewed by the LIKE A KING INC. team. LIKE A KING INC. is a software studio in New York.