Ring-1T: The Trillion-Parameter Thinking Model from Ant Group

1. Introduction

Just ten days after releasing Ling-1T, Ant Group made headlines again by unveiling Ring-1T — the world’s first open-source trillion-parameter thinking model.

Unlike Ling-1T, which focuses on general language understanding, Ring-1T is designed specifically for deep reasoning and logical thinking. In fact, during internal tests on the International Mathematical Olympiad (IMO) benchmark, Ring-1T achieved silver-medal performance, marking a huge leap for open-source AI.

In this article, we’ll walk through what makes Ring-1T unique, its performance highlights, and — most importantly — how you can install, run, and experiment with this groundbreaking “thinking” model.

2. What Is Ring-1T?

Ring-1T is an open-source “thinking-type” large language model developed by Ant Group’s Bailing AI team.
It’s built on a trillion-parameter Mixture-of-Experts (MoE) architecture — meaning only a small subset of experts (about 50 B active parameters) are activated for each token, dramatically improving inference efficiency.

The model goes beyond standard LLMs by incorporating reasoning-focused reinforcement learning:

• RLVR (Reinforcement Learning from Verifiable Rewards): trains the model to verify its reasoning chain.
• RLHF (Reinforcement Learning from Human Feedback): balances reasoning accuracy with natural language fluency.

These techniques are powered by Ant’s self-developed systems:

• IcePop Algorithm (“Popsicle”) — filters unstable gradients to stabilize trillion-scale training.
• ASystem RL Framework — handles distributed GPU scheduling, reward evaluation, and serverless sandbox execution.

🔗 Model Pages

• HuggingFace: inclusionAI/Ring-1T
• ModelScope: inclusionAI/Ring-1T

3. Key Highlights

1. Trillion-Parameter MoE Design

   • Uses sparse activation to maintain high efficiency while achieving unprecedented reasoning capacity.

2. Reinforcement-Enhanced Thinking Ability

   • RLVR enables multi-step reasoning with verifiable logic chains.

3. World-Class Math and Coding Skills

   • Achieved IMO silver medal level in mathematics; solved 5/6 problems in the ICPC World Finals 2025.

4. Stable Large-Scale RL via IcePop

   • Monitors training “temperature” to prevent gradient explosion or collapse.

5. ASystem Infrastructure

   • Supports trillion-scale distributed reinforcement learning with GPU peer-to-peer synchronization.

4. How to Install and Use Ring-1T

You can easily try Ring-1T using HuggingFace, ModelScope, or OpenRouter APIs.

Method 1 — HuggingFace + Transformers

pip install transformers accelerate vllm

from transformers import AutoTokenizer, AutoModelForCausalLM

model_name = "inclusionAI/Ring-1T"

tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name, device_map="auto", torch_dtype="auto")

prompt = "Prove: if f(x) is an odd function and f(x)+f(1-x)=0, find f(1/2)"
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=512)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Method 2 — OpenRouter API

curl https://openrouter.ai/api/v1/chat/completions \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "inclusionAI/ring-1t",
    "messages": [{"role": "user", "content": "Write a simple Flappy Bird game in Python"}]
  }'

Method 3 — ModelScope Local Inference

pip install modelscope
from modelscope import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("inclusionAI/Ring-1T", device_map="auto")
tokenizer = AutoTokenizer.from_pretrained("inclusionAI/Ring-1T")

inputs = tokenizer("Generate a simple Snake game with scoring and pause function", return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=300)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

5. Performance Overview

Model	Parameters	Reasoning (ARC-AGI-v1)	Math (AIME25)	Code (LiveCodeBench)	Overall (Arena-Hard-v2)
Ring-1T	1 T	Open-source SOTA	≈ IMO Silver Medal	Strong	2nd, just below GPT-5-Thinking
Ling-1T	1 T	Good	Moderate	Good	3rd tier
DeepSeek-V3.1	671 B	High	High	High	2nd tier
Gemini-2.5-Pro	N/A	High	High	High	Closed-source top
GPT-5-Thinking (High)	N/A	Best	Best	Best	1st tier

6. Usage Tips

1. Use Reasoning Prompts
   Add instructions like “Think step by step” or “Verify each reasoning step” to get more accurate answers.

2. Leverage Long Contexts
   Supports up to 128 K tokens, ideal for academic papers, codebases, or reasoning datasets.

3. Language Support
   Works best in English and Chinese. Multilingual reasoning still under optimization.

4. Hardware Recommendation

• Minimum: 48 GB GPU memory (A100/H100/L40S)

• Use FP8 or bfloat16 inference for efficiency.

7. Real-World Testing

During testing, Ring-1T successfully generated a playable Flappy Bird clone in HTML and JavaScript — complete with animation, collision detection, and scoring logic.

For reasoning tasks, the model demonstrates a human-like reasoning flow: outlining assumptions, validating logic, and providing structured explanations.

In creative writing, Ring-1T generated a podcast script about Su Shi and Zhang Dun’s political feud, blending historical accuracy with storytelling flair — even including suggested sound effects.

8. Editor’s Notes

If you’re a:

• Researcher → Ring-1T offers a goldmine for studying reasoning-enhanced RLHF.

• Developer → Great for building educational, logic, or math-related apps.

• Content Creator → Reliable for long-form storytelling and interactive writing.

Ring-1T proves that open-source models can now compete directly with closed-source giants like GPT-5.

Though minor issues like language mix and self-consistency remain, its trillion-scale architecture and open accessibility mark a new milestone for the open AI ecosystem.

9. FAQ

Q1: What’s the difference between Ring-1T and Ling-1T?

A: Ling-1T is a general-purpose LLM, while Ring-1T is optimized for deep reasoning, logical analysis, and mathematical problem-solving.

Q2: What hardware do I need to run it?

A: At least 48 GB of GPU memory (A100/H100/L40S). Distributed inference with vLLM or DeepSpeed is recommended for lower VRAM systems.

Q3: Can I run it on CPU?

A: Technically yes, but it’s extremely slow. You’re better off running it on a GPU cloud (e.g., LightNode, AWS, or RunPod).

Q4: Does it support Chinese?

A: Yes. Ring-1T supports both English and Chinese input, with highly stable reasoning performance in Chinese.

Q5: Is it fully open-source?

A: Yes — the model weights and configs are available under an open license on HuggingFace and ModelScope.

Q6: How to get more accurate reasoning outputs?

A: Add structured reasoning prompts like “Explain your reasoning step by step” or “Validate each step logically.”

Q7: Can it be used commercially?

A: Yes. The model is open-source and can be used for academic, research, and commercial purposes under the stated license.