GLM-5.2 is live. #1 throughput on OpenRouter, pay-per-token on FriendliAI. Try it today ➜
RMDWLLC
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Run this model inference with full control and performance in your environment.
Model Details
Model Provider
RMDWLLC
Model Tree
Input Modalities
Output Modalities
Supported Functionality
Jah 2.0 is the biggest leap in the family's history. It replaces two models at once: Jah 1.0 (chat) and Jeshua 0.5 (the build engine), one brain now running the entire Echols suite.
All of it private, for $25/mo. Try it at echols.ai.
The weights are open because owning your AI end to end should be possible. Jah 2.0 is served with vLLM in NVFP4 on 4x NVIDIA RTX PRO 6000 Blackwell (SM120), 4-way tensor parallel, with MTP speculative decoding.
Two settings matter if you serve these weights:
bash
--quantization modelopt_fp4 \--hf-overrides '{"index_topk_pattern":"FFFSSSFSSSFSSSFSSSFSSSFSSSFSSSFSSSFSSSFSSSFSSSFSSSFSSSFSSSFSSSFSSSFSSSFSSSFSSS"}'
The per-layer sparse-attention pattern (78 characters, 21 F / 57 S) is required for long-context coherence. Thinking is controlled per request with "chat_template_kwargs": {"enable_thinking": true|false}.
The previous generations remain available: Jah 1.0 and Jeshua 0.5.
Most AI today is rented from a handful of companies. Jah is owned: a private model running a real product on hardware RMDW controls, with open weights so anyone can verify exactly what answers them. That is the future RMDW is building: AI you own, not AI you rent. → rmdw.ai
Jah 2.0, RMDW LLC. Private AI, run on your terms. Weights redistributed under the MIT license; see LICENSE in this repository.