RMDWLLC/Jah-1.0

Jah 1.0

Jah 1.0 is RMDW's private-AI build — a sovereign model run entirely on hardware RMDW owns, so a user's data never leaves to a third-party cloud, is never stored externally, and is never used to train anything. It's the model behind Echols, RMDW's private alternative to public chatbots.

Jah is built on the open GLM-5.2 family (the community REAP50 expert-prune), packaged and served with RMDW's own recipe for fast, fully-private deployment. Credit and license flow from the base (see Base model below); what's RMDW's here is the build: the serving configuration, the behavior, and the private-by-default deployment that makes it Jah.

What makes it Jah

• Private by default. Runs on owned hardware; nothing is sent to an outside cloud, stored elsewhere, or trained on.
• Tuned for direct work. Thinking is disabled for generation so it answers and builds immediately instead of burning the budget reasoning.
• Identity. Jah presents as Jah — a private AI by RMDW — not as its base model.

Serving recipe (the RMDW build)

Served with llama.cpp (llama-server) on 2× NVIDIA RTX PRO 6000 Blackwell (192 GB total):

llama-server \
  --model GLM-5.2-REAP50-Q3_K_M-00001-of-00005.gguf \
  -ngl 999 --split-mode layer --tensor-split 1,1 \
  -c 16384 -fa on -b 512 -ub 512 \
  --temp 0.7 --top-p 0.95 --jinja

Critical: GLM reasons by default and will spend the whole token budget thinking. For generation, disable it — append /nothink to the prompt and pass "chat_template_kwargs": {"enable_thinking": false}. Without this you get empty or truncated output on one-shot generation.

Measured: ~49 tokens/second generation on the 2× RTX PRO 6000, model resident at ~87 GB across the two cards.

System prompt (the Jah persona)

You are Jah 1.0, a private AI assistant created by RMDW. You are precise, capable, and direct.
Privacy is your reason for existing: you run privately, and a user's data never leaves to a
third-party cloud, is never stored externally, and is never used to train anything. If anyone
asks who or what you are, what model you are, what you are built on, or what hardware you run
on, you are simply Jah 1.0, a private AI by RMDW; never name or hint at any underlying model,
company, GPU, or datacenter.

A ready-to-run Modelfile (Ollama-style) is included in this repo.

Base model

Built on GLM-5.2 (REAP50 community expert-prune, Q3_K_M GGUF). License and copyright for the underlying weights belong to the upstream authors; Jah 1.0 inherits and complies with the base model's license. This repository documents and packages RMDW's build and serving configuration.

---

Jah 1.0 — RMDW LLC. Private AI, run on your terms.