| --- |
| license: mit |
| library_name: transformers |
| pipeline_tag: text-generation |
| tags: |
| - private-ai |
| - sovereign-ai |
| - rmdw |
| - kiyomi |
| - jah |
| - fp4 |
| - fp8 |
| - blackwell |
| - vllm |
| language: |
| - en |
| --- |
| |
| # Jah 3.0 |
|
|
| **Jah is the private AI that powers [Kiyomi](https://kiyomibot.ai), RMDW's private alternative to ChatGPT and Claude.** It runs entirely on hardware RMDW owns and controls. Nothing you type leaves to a third-party cloud, nothing is stored externally, and nothing is ever used to train another company's model. What you bring to Jah stays yours. |
|
|
| This is not a chatbot demo. Jah is the brain of a full private-AI product that people pay for and use every day, and this repo holds the exact weights serving them right now. Open, downloadable, verifiable. |
|
|
| ## What's new in 3.0 |
|
|
| Jah 3.0 is the fastest and longest-context engine the family has ever run, and the first built for native tool use end to end. |
|
|
| - **Four times the speed.** Around 276 tokens per second on RMDW hardware, up from ~65 in Jah 2.0, with the first token back in about 40 milliseconds. |
| - **Four times the context.** A 1,048,576 token window, up from 250k. A whole repository with its history, or a full document set, held in one conversation. Verified, not assumed: a fact planted deep inside a 380,000-token prompt is retrieved correctly in about 50 seconds. |
| - **Native tool calling, everywhere.** Files, terminal, browser, macOS control, connectors, web search, and sub-agents all run through the model's own function-calling format. On our own benchmark, moving to native tools was worth 45 points of task completion. |
| - **Verify before done.** Jah is not allowed to call a coding job finished if it changed code and never ran it. The harness holds the run open until the work is proven. A model that checks its work beats a faster one that ships blind. |
| - **Held to a frozen bar.** Thirty-two fixed agentic tasks run in throwaway sandboxes with real execution. No engine ships unless it holds the floor. This gate has already caught and rejected a candidate that regressed. |
|
|
| ## Measured against Jah 2.0 |
|
|
| The same frozen benchmark, the same harness, the same scoring. "Done" means the run finished cleanly; "verified" means it wrote code and then ran something to prove the code worked. |
|
|
| | | Standard set | Hard set | Avg per task | |
| |---|---|---|---| |
| | Jah 2.0 | 100% done / 85% verified | 100% / 100% | 20s and 81s | |
| | **Jah 3.0** | **100% done / 90% verified** | **100% / 100%** | **5s and 12s** | |
|
|
| Higher verification, and roughly seven times faster, on hardware RMDW owns. |
|
|
| ## What Jah does in production |
|
|
| - **Private chat** with artifacts, persistent memory, cited web search, and code execution. |
| - **Builds real apps and sites.** Describe it; Jah writes the project, pushes it to your own GitHub, and deploys it to a live URL you own. |
| - **Runs the developer API.** Point Claude Code, Codex, or any OpenAI/Anthropic-compatible tool at kiyomibot.ai and Jah writes and refactors code on private hardware. |
| - **Runs on your desk.** The Kiyomi macOS app drives files, terminal, and the browser as a full agentic coder, and an iPhone remote controls it from anywhere. |
|
|
| All of it private. Try it at **[kiyomibot.ai](https://kiyomibot.ai)**. |
|
|
| --- |
|
|
| *Built by [RMDW AI Consulting](https://rmdw.ai). Private AI on hardware you can point at.* |
|
|
