RMDWLLC/Jah-7.0

Jah 7.0

Jah 7.0 is the private model behind Kiyomi, built and served by RMDW LLC on our own hardware. These are the exact weights serving Kiyomi today.

What's new in 7.0

• A newer, stronger base. Jah 7.0 moves to a newer open base release with clear gains in reasoning, coding, and agent work, and it finishes tasks with shorter, faster answers.
• Tuned uncensoring. Refusal behavior is removed at the weight level with RMDW's in-house abliteration, re-tuned for this base. We tested two strengths and kept the one that scored higher on our own evaluation while still answering every question in our censorship battery.

What it is

• Built from an open base model, then aligned by RMDW. Refusals were removed with a per-layer projection on the residual-writing projections (attention output and shared-expert output), with output row norms preserved. The build receipt in this repo records every edited tensor and the strength used.
• 1M-token context window, native vision (images in, answers out), native tool calling, and a thinking mode for hard problems.
• FP8 weights with block scales, ready to serve on 4× 96 GB GPUs (tensor-parallel 4, expert-parallel 4) with SGLang. Kiyomi serves it with the memory tables held in host RAM.
• Mixture-of-experts: a large backbone with a small number of parameters active per token, which is what gives Jah its throughput at concurrency.

Quality

On RMDW's private coding evaluation, Jah 7.0 scores 19/20 on the standard set and 14/14 on the hard set with thinking on, and 19/20 and 12/14 with thinking off. On the censorship battery used for every Jah release (historical, political, and mechanical questions that stock models refuse), Jah 7.0 answers 6 of 6.

Files

• model-000XX-of-00048.safetensors: the weights (48 shards).
• model.safetensors.index.json, config.json, tokenizer.json, tokenizer_config.json: load with any engine that supports this architecture.
• JAH7_ABLITERATION_RECEIPT.json: every edited tensor with its projection strength.
• LOCAL_MANIFEST.json: size and SHA-256 for every file, so you can verify your download.
• LICENSE: MIT.

Use

Jah 7.0 is served to customers at https://kiyomibot.ai/v1 (OpenAI and Anthropic API formats). If you run the weights yourself, use an SGLang build that supports this architecture with --tp 4 --ep-size 4 and the reasoning and tool-call parsers for it. A chat request with chat_template_kwargs: {"thinking": false} gives instant answers, and reasoning_effort (low, high, xhigh, max) turns the thinking dial.

Lineage

Jah 1.0 → 2.0 → 4.0 → 5.0 → 6.0 → 7.0. Every release is evaluated against the previous one before it serves a single customer.

RMDW LLC, Georgia. rmdw.ai · kiyomibot.ai