---
base_model: Qwen/Qwen3.5-9B
library_name: transformers
pipeline_tag: image-text-to-text
tags:
- qwen3.5
- qwen
- rrchat
- viper
- reasoning
- math
- code
- agent
- tool-use
- function-calling
- russian
- unsloth
license: other
language:
- ru
- en
datasets:
- open-r1/OpenR1-Math-220k
- nvidia/Nemotron-SFT-Agentic-v2
- Vikhrmodels/GrandMaster-PRO-MAX
- nvidia/OpenMathInstruct-2
- nvidia/OpenCodeInstruct
- m-a-p/CodeFeedback-Filtered-Instruction
---

<div align="center">
  <img src="viperbanner.png" alt="Viper-9B Banner" width="100%" style="border-radius: 10px; margin-bottom: 20px;">
</div>

# Viper-9B

**rrchat/Viper-9B** is the flagship agentic, coding, and reasoning language model of the **RR-Gen1** lineup developed by **rrchat** ([huggingface.co/rrchat](https://huggingface.co/rrchat)), fine-tuned on top of **Qwen3.5-9B**.

Viper-9B is engineered for autonomous tool orchestration, structured function calling, advanced code synthesis, competitive programming, and robust bilingual Russian–English instruction following.

Looking for quantized GGUF weights for llama.cpp / LM Studio / Ollama?  
👉 **[`rrchat/Viper-9B-GGUF`](https://huggingface.co/rrchat/Viper-9B-GGUF)**

---

## 🌟 Key Highlights & Capabilities

* **Flagship RR-Gen1 Model:** The leading 9B parameter model of the RR-Gen1 family, optimized for high-complexity agentic workflows, competitive programming, and bilingual problem solving.
* **Autonomous Tool Use & Function Calling:** Native support for structured XML function execution (`<tool_call>...<tool_response>`) preserving near-perfect formatting and AST execution accuracy (57.0% BFCL AST match, 99.0% valid call rate).
* **Enhanced Competitive Coding:** Exceptional leap in algorithmic problem solving on **LiveCodeBench v6** (**33.0%** vs Base 19.0%, with **+18.2%** on Easy and **+14.7%** on Medium problems).
* **Python Function Synthesis:** Achieves **84.7%** on **MBPP+** (+1.4% gain over base model).
* **Bilingual Russian–English Math:** Delivers **44.4%** on Russian mathematical reasoning (**MGSM ru**), outperforming the base foundation model (+2.4%).
* **General Knowledge Retention:** Retains broad academic, STEM, and humanities knowledge with **74.0%** accuracy on **MMLU-Redux**.
* **Calibrated Honesty & R-Tuning:** Trained with R-Tuning to reduce confident hallucinations on unknown topics while maintaining high factuality.

---

## 🧪 Benchmark Highlights (vs. Base Model)

All benchmarks were evaluated locally on an identical test harness under strictly identical conditions:

| Benchmark | Mode / Harness | Base (`Qwen3.5-9B`) | Viper-9B | Delta vs. Base | Domain / Capability |
| :--- | :---: | :---: | :---: | :---: | :--- |
| **LiveCodeBench v6** | **Deterministic Pass@1** | 19.0% | **33.0%** | **+14.0%** | Competitive programming & algorithmic tasks |
| ↳ *LCB v6 Easy* | Deterministic Pass@1 | 42.4% | **60.6%** | **+18.2%** | Standard algorithmic challenges |
| ↳ *LCB v6 Medium* | Deterministic Pass@1 | 14.7% | **29.4%** | **+14.7%** | Complex dynamic programming & data structures |
| ↳ *LCB v6 Hard* | Deterministic Pass@1 | 0.0% | **9.1%** | **+9.1%** | Advanced competitive problem solving |
| **MBPP+ (seed 42)** | **EvalPlus Pass@1** | 83.3% | **84.7%** | **+1.4%** | Python function synthesis & unit tests |
| **MGSM ru (Math)** | **Numeric Match** | 42.0% | **44.4%** | **+2.4%** | Russian multi-step mathematical reasoning |
| **MMLU-Redux** | **Choice Accuracy** | 73.7% | **74.0%** | **+0.3%** | Broad multidisciplinary STEM & humanities |
| **BFCL v4 (Tool Use)** | **AST Function Match** | 58.0% | **57.0%** | **Parity (~57%)** | Berkeley Function Calling / Agentic Tool Calls |
| **IFEval (Instruction)** | **Strict Compliance** | 56.6% | **54.7%** | **Parity (~55%)** | Complex constraint & prompt compliance |

> [!NOTE]
> **Disclaimer on Evaluations:**  
> All benchmarks presented above were conducted **manually and independently** on local test suites under controlled, strictly identical conditions (greedy non-thinking mode, max_tokens 1024, llama-server with jinja chat formatting). The results reflect real-world local execution performance and may differ from theoretical, thinking-mode, or officially reported figures published on the official base model card.

---

## 💻 Quickstart (Transformers / Python)

```python
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "rrchat/Viper-9B"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto"
)

messages = [
    {"role": "user", "content": "Привет! Расскажи о себе и помоги решить задачу."}
]

prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)

with torch.no_grad():
    outputs = model.generate(**inputs, max_new_tokens=512, temperature=0.7, top_p=0.8)

print(tokenizer.decode(outputs[0][inputs.input_ids.shape[1]:], skip_special_tokens=True))
```

---

## 🛠️ Tool Calling Format

Viper-9B utilizes structured ChatML syntax for function calling:

```xml
<|im_start|>system
# Tools

You have access to the following functions:

<tools>
{"name": "get_weather", "description": "Get current weather", "parameters": {"type": "object", "properties": {"location": {"type": "string"}}, "required": ["location"]}}
</tools>
<|im_end|>
<|im_start|>user
Какая погода в Москве?<|im_end|>
<|im_start|>assistant
<tool_call>
<function=get_weather>
<parameter=location>
Москва
</parameter>
</function>
</tool_call><|im_end|>
```