File size: 2,900 Bytes
7c6b0a0 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 | ---
license: mit
library_name: sglang
pipeline_tag: image-text-to-text
tags:
- rmdw
- jah
- private-ai
- fp8
- mixture-of-experts
- vision
---
# Jah 5.0
Jah 5.0 is the model behind [Kiyomi](https://kiyomibot.ai), RMDW's private AI platform. These are the exact weights serving Kiyomi customers right now. Open, downloadable, verifiable. Don't take our word for what powers your work: inspect it.
Same promise as every Jah release: open weights, no secrets.
## What's new in 5.0
- **A 12x jump in knowledge capacity over Jah 4.0.** 321B parameters with 18B active per token. Only a fraction of the model wakes for each token, so it streams fast for its size.
- **Thinking on demand.** Ask for deep reasoning and it thinks before it answers, in a cleanly separated reasoning channel. Ask plain and it answers instantly.
- **Perfect run on our hardest internal suite** with the orchestrated engine, and it solves problems single-shot that Jah 4.0 could not solve at any setting.
- **Vision.** Send a screenshot, an invoice, a whiteboard photo. Jah reads it like text.
- **Native 1M token context.** 1,048,576 positions, no scaling tricks.
## Specs
| | |
|---|---|
| Parameters | 321B total, 18B active per token |
| Attention | Hybrid: linear attention + sparse attention, 45 layers |
| Context | 1,048,576 tokens native |
| Precision | FP8 (block quantized), BF16 sensitive paths |
| Modalities | Text + vision in, text out |
| Tool calling | Native |
| License | MIT |
## Running it
Full precision wants roughly 330GB of GPU memory plus KV cache. A 4x RTX PRO 6000 class rig (384GB) serves it comfortably; that is exactly what Kiyomi runs on. Community 4-bit quantizations bring it to a 512GB Mac Studio.
Serve with SGLang (a current build with this architecture's day-0 support):
```bash
python3 -m sglang.launch_server \
--model-path RMDWLLC/Jah-5.0 \
--tp 4 --mem-fraction-static 0.85 \
--context-length 1048576 \
--kv-cache-dtype bfloat16 \
--reasoning-parser glm45 --tool-call-parser glm47 \
--enable-multimodal --trust-remote-code
```
On workstation Blackwell GPUs (RTX PRO 6000 class, SM120) add `--attention-backend triton` and disable DeepGEMM (`SGLANG_ENABLE_JIT_DEEPGEMM=0 SGLANG_ENABLE_DEEP_GEMM=0 SGLANG_OPT_DEEPGEMM_HC_PRENORM=false`); datacenter kernels for this architecture have not landed for SM120 yet.
Recommended sampling: `temperature=1.0, top_p=0.95` with thinking on; `temperature=0.6` for instant mode.
**Thinking off** (instant mode): pre-close the reasoning block by prefilling the assistant turn with `<think></think>` and continuing the final message. **Thinking on** is the default.
## No rack at home?
That's the API: the same Jah 5.0, on our hardware, one line away. Keys at [kiyomibot.ai](https://kiyomibot.ai).
---
Built and served by [RMDW](https://rmdw.ai), Atlanta. Private by physics: your data never leaves the machine.
|