README.md

---
license: mit
language:
- en
base_model:
- inclusionAI/Ling-mini-base-2.0-20T
pipeline_tag: text-generation
library_name: transformers
tags:
- moe
---

# Ring-mini-linear-2.0

<p align="center">
    <img src="https://mdn.alipayobjects.com/huamei_qa8qxu/afts/img/A*4QxcQrBlTiAAAAAAQXAAAAgAemJ7AQ/original" width="100"/>
<p>
<p align="center">🤗 <a href="https://huggingface.co/inclusionAI">Hugging Face</a>&nbsp&nbsp | &nbsp&nbsp🤖 <a href="https://modelscope.cn/organization/inclusionAI">ModelScope</a></p>

## Introduction

We are excited to announce the official open-source release of Ring-mini-linear-2.0!

Building on the success of our Ling 2.0 series, this model continues to leverage a powerful hybrid architecture of linear and standard attention, perfectly balancing high performance with superior efficiency. By integrating our proven MoE design with optimizations like a 1/32 expert activation ratio and MTP layers, Ring-mini-linear achieves the performance of a 8 B dense model while activating only 1.4 B parameters.  This model was converted from [Ling-mini-base-2.0](https://huggingface.co/inclusionAI/Ling-mini-base-2.0-20T), further trained on an additional 600 B tokens.
When it comes to benchmarks, Ring-mini-linear-2.0 not only holds its own against standard attention models (like ring-mini-2) but also outperforms other open-source MoE and Dense models in its class on several demanding tasks. Plus, with native support for a 128k long context, it's faster and more precise than ever, especially when handling long-form inputs and outputs.

<div style="display: flex; justify-content: center;">
  <div style="text-align: center;">
    <img src="https://cdn-uploads.huggingface.co/production/uploads/68d20104a6f8ea66da0cb447/UsAtWWsWB9eXcMxV5iCCa.png" width="600">
    <p style="margin-top: 8px; font-size: 14px;"><strong>Figure 1:</strong> Hybrid Linear Model Architecture</p>
  </div>
</div>

## Evaluation
To properly evaluate the model's reasoning capabilities, we compared it against 3 other models—Ring-mini-2.0, Qwen3-8B-thinking, and GPT-OSS-20B-Medium—on 6 challenging reasoning benchmarks spanning mathematics, coding, and science. The results demonstrate that the performance of the hybrid linear architecture is by no means inferior to that of standard softmax attention; in fact, it even outperforms the other models on 3 of the benchmarks.
<div style="display: flex; justify-content: center;">
  <div style="text-align: center;">
    <img src="https://cdn-uploads.huggingface.co/production/uploads/68d20104a6f8ea66da0cb447/_tjjgBEBlankfrWUY0N9i.png" width="800">
    <p style="margin-top: 8px; font-size: 14px;"><strong>Figure 2:</strong> Model Performance Comparison </p>
  </div>
</div>

Here is a demo of a small Snake game, with the code generated by our model.
<div style="display: flex; justify-content: center;">
  <div style="text-align: center;">
    <img src="https://mdn.alipayobjects.com/huamei_jcuiuk/afts/img/tqfCQoTqRdAAAAAAgZAAAAgADr6CAQFr/original" width="800">
    <p style="margin-top: 8px; font-size: 14px;"><strong>Figure 3:</strong> Snake Game </p>
  </div>
</div>


## Linear Attention, Highly Sparse，High-Speed Generation

Thanks to its hybrid attention mechanism and highly sparse MoE architecture, Ring-mini-linear-2.0 achieves near-linear time complexity and constant space complexity, resulting in outstanding inference efficiency. To fully demonstrate this advantage, we conducted a head-to-head comparison between our model and top-tier competitors of similar size or performance.
The results are remarkable. In the prefill stage, Ring-mini-linear-2.0's performance is exceptional; when the context length exceeds 256k, its throughput is over 12 times higher than that of Qwen3-8B. Furthermore, in the high-concurrency decode stage, its capabilities are even more pronounced. For generation lengths exceeding 32k, its throughput easily surpasses 12 times that of Qwen3-8B.

<div style="display: flex; justify-content: center; align-items: flex-start; gap: 20px;">
  <div style="text-align: center;">
    <img src="https://cdn-uploads.huggingface.co/production/uploads/68d20104a6f8ea66da0cb447/O9gHLIOCdpWvBbPC6bMM5.webp" width="500">
    <p style="margin-top: 8px; font-size: 14px;"><strong>Figure 4:</strong> Ring-mini-linear-2.0 prefill throughput</p>
  </div>
  
  <div style="text-align: center;">
    <p align="center">
      <img src="https://cdn-uploads.huggingface.co/production/uploads/68d20104a6f8ea66da0cb447/AvMTStWFX-Frzv-vOzyr6.webp" width="500">
    </p>
    <p style="margin-top: 8px; font-size: 14px;"><strong>Figure 5:</strong> Ring-mini-linear-2.0 decode throughput</p>
  </div>

</div>


## Model Downloads

<div align="center">

|     **Model**      | **#Total Params** | **#Activated Params** | **Context Length** | **Download** |
| :----------------: | :---------------: | :-------------------: | :----------------: | :----------: |
| Ring-mini-linear-2.0 |       16.8B       |         1.4B         |        128K         |      [🤗 HuggingFace](https://huggingface.co/inclusionAI/Ring-mini-linear-2.0) <br>[🤖 Modelscope](https://modelscope.cn/models/inclusionAI/Ring-mini-linear-2.0)|
</div>

## Quickstart

### Requirements

```bash
pip install flash-linear-attention==0.3.2
pip install transformers==4.56.1
```

### 🤗 Hugging Face Transformers

```python
from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "inclusionAI/Ring-mini-linear-2.0"

model = AutoModelForCausalLM.from_pretrained(
    model_name,
    dtype="auto",
    device_map="auto",
    trust_remote_code=True,
)
tokenizer = AutoTokenizer.from_pretrained(model_name)


prompts = [
    "Give me a short introduction to large language models."
]
input_texts = []
for prompt in prompts:
    messages = [
        {"role": "user", "content": prompt}
    ]
    text = tokenizer.apply_chat_template(
        messages,
        tokenize=False,
        add_generation_prompt=True,
        enable_thinking=True
    )
    input_texts.append(text)

print(input_texts)

model_inputs = tokenizer(input_texts, return_tensors="pt", return_token_type_ids=False, padding=True, padding_side='left').to(model.device)

generated_ids = model.generate(
    **model_inputs,
    max_new_tokens=8192,
    do_sample=False,
)
generated_ids = [
    output_ids[len(input_ids):] for input_ids, output_ids in zip(model_inputs.input_ids, generated_ids)
]

responses = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)

print("*" * 30)
print(responses)
print("*" * 30)
```

### SGLang
```bash
python -m sglang.launch_server \
    --model-path <model_path> \
    --trust-remote-code \
    --tp-size 1 \
    --disable-radix-cache \
    --json-model-override-args "{\"linear_backend\": \"seg_la\"}"
```
### vLLM
 
## Citation