# rrchat AI
**Open-source compact reasoning models & private mobile-ready local AI**
[](https://github.com/leftik1/rrchat)
[](https://huggingface.co/rrchat)
[](https://github.com/ggerganov/llama.cpp)
[](https://www.apache.org/licenses/LICENSE-2.0)
---
### ๐ Overview
**rrchat AI** is an open-source initiative focused on high-efficiency, private local AI: combining compact bilingual (RU/EN) reasoning models with native mobile interfaces.
Our goal is to bring sovereign intelligence directly into your workflow โ zero cloud telemetry, maximum throughput on consumer hardware, and seamless orchestration between mobile devices and local GPU workstations.
---
### ๐ง Models: RR-Gen1 Series
The **RR-Gen1** lineup spans from ultra-fast 1.7B edge assistants to a 9B agentic flagship, optimized for agentic tool use, mathematical reasoning, and competitive coding:
| Model | Size / Active | Base Architecture | Primary Strengths | Recommended VRAM | Weights & GGUF |
| :--- | :---: | :---: | :--- | :---: | :---: |
| **Viper-9B** *(Flagship)* | 9B | Qwen3.5-9B | Competitive Coding (LCB 33%), Agentic Tool Calling, Math CoT, R-Tuning Honesty | ~8โ12 GB | [Base](https://huggingface.co/rrchat/Viper-9B) ยท [GGUF](https://huggingface.co/rrchat/Viper-9B-GGUF) |
| **Cerberus-E4B** | 4B (7.5B Total) | Gemma 4-E4B | Balanced Multi-Task Reasoning, Native Gemma Tool Schemas, Multimodal | ~4โ8 GB | [Base](https://huggingface.co/rrchat/Cerberus-E4B) ยท [GGUF](https://huggingface.co/rrchat/Cerberus-E4B-GGUF) |
| **Mercury-1.7B** | 1.7B | LFM3-VL-2B | Extreme Speed (~180โ200 t/s), Linear Attention + Short-Conv, SigLIP Vision | ~2โ4 GB | [Base](https://huggingface.co/rrchat/Mercury-1.7B) ยท [GGUF](https://huggingface.co/rrchat/Mercury-1.7B-GGUF) |
| **Mercury-Acad-1.7B** | 1.7B | LFM3-VL-2B | Academic & Scientific Reasoning, Mathematical CoT, STEM Domain Specialist | ~2โ4 GB | [Base](https://huggingface.co/rrchat/Mercury-Acad-1.7B) ยท [GGUF](https://huggingface.co/rrchat/Mercury-Acad-1.7B-GGUF) |
#### ๐ฌ Lineup Highlights
* ๐ **Viper-9B**: The flagship for software engineering and complex reasoning. Demonstrates +14% absolute gain on LiveCodeBench v6, 84.7% on MBPP+, and calibrated factual honesty without confident hallucinations.
* ๐บ **Cerberus-E4B**: Efficient mid-weight specialist built on Google's Gemma 4 backbone, excelling in multi-turn tool interaction and structured bilingual reasoning.
* โก **Mercury-1.7B**: Sub-2B powerhouse powered by Liquid AI's hybrid linear attention and causal depthwise convolutions, delivering desktop-class reasoning at up to 200 tokens/sec on edge hardware.
* ๐ **Mercury-Acad-1.7B**: Tailored for mathematical research, paper comprehension, and STEM problem solving with dual-mode explicit reasoning (`