Ring-1T Now on SiliconFlow: The World's First Open-Source Trillion-Parameter Thinking Model

TL;DR: Ring-1T — the world's first open-source trillion-parameter thinking model — is now live on SiliconFlow. Built on Ling 2.0 architecture (1T total / 50B active parameters, 131K context), it introduces the Icepop algorithm and ASystem for stable large-scale reinforcement learning across MoE models. Achieving silver-medal-level performance at IMO 2025, Ring-1T sets a new benchmark for open-source deep reasoning.

Now, start building with Ring-1T today through SiliconFlow's production-ready API!

Today, we're excited to bring Ring-1T to SiliconFlow — the first open-source trillion-parameter thinking model developed by Ant Group's inclusionAI team. Following the success of Ling-1T, Ring-1T extends the Ling 2.0 architecture to trillion scale, combining dense-level reasoning quality with efficient MoE activation. With the newly introduced Icepop algorithm and ASystem for stable large-scale reinforcement learning, it pushes open-source deep reasoning to a new frontier — from mathematics and programming to complex natural-language inference.

With SiliconFlow's Ring-1T API, you can expect:

• Competitive Pricing: $0.57/M tokens (input) and $2.28/M tokens (output).

• Trillion-Parameter MoE Architecture: 1T total parameters with ~50B active per token, delivering dense-level reasoning power with efficient resource utilization.

• Stable Reinforcement Learning: Powered by the Icepop algorithm and ASystem, enabling smooth, long-term RL training across trillion-scale MoE models.

• Advanced Reasoning & Math: Achieves silver-medal-level performance at IMO 2025, demonstrating world-class mathematical and logical reasoning ability.

• 131K Context Window: Supports long-context document analysis and multi-turn reasoning with stable performance and low latency.

Whether you're developing analytical systems, educational applications, or reasoning-based agents, you can now access Ring-1T directly through SiliconFlow's API to power your next-generation solutions.

Why Ring-1T Matters

Evolved Deep Reasoning Capabilities

Developed by Ant Group's inclusionAI team, Ring-1T represents a new milestone in trillion-scale reasoning.

Compared to its preview version, it delivers more balanced and robust performance across mathematics, programming, and logical reasoning — achieving state-of-the-art open-source results on leading benchmarks including AIME 25, HMMT 25, LiveCodeBench, CodeForce, ARC-AGI-1, and Arena-Hard-v2.0.

To ensure fairness, Ring-1T underwent string and semantic decontamination throughout pre-training, fine-tuning, and reinforcement-learning stages.

Its results on IMO 2025 and ICPC World Finals 2025 further validate its deep reasoning strength, with the model reaching silver-medal-level mathematical performance and solving five of twelve ICPC problems — rivaling top closed-source systems like Gemini-2.5-Pro and GPT-5-Thinking.

To see how Ring-1T handles mathematical reasoning in practice, we gave it a logic-based coordinate problem. It broke down the conditions step by step, explained its reasoning clearly, and reached the correct answer with full consistency — a concise example of its structured and transparent thought process.

Prompt and reasoning demo below:

Let A={1,2,3,4,5,6,7,8}A = \{1, 2, 3, 4, 5, 6, 7, 8\}A={1,2,3,4,5,6,7,8}, and
M={(xi,yi)∣xi∈A, yi∈A}M = \{(x_i, y_i) \mid x_i \in A,\, y_i \in A\}M={(xi,yi)∣xi∈A,yi∈A}.
Select nnn ordered pairs from MMM to form a sequence:
(x1,y1),…,(xn,yn)(x_1, y_1), \ldots, (x_n, y_n)(x1,y1),…,(xn,yn)
where each pair of adjacent terms (xi,yi)(x_i, y_i)(xi,yi) and (xi+1,yi+1)(x_{i+1}, y_{i+1})(xi+1,yi+1) satisfies either
{∣xi+1−xi∣=3∣yi+1−yi∣=4or{∣xi+1−xi∣=4∣yi+1−yi∣=3\begin{cases} |x_{i+1} - x_i| = 3 \\ |y_{i+1} - y_i| = 4 \end{cases} \quad \text{or} \quad \begin{cases} |x_{i+1} - x_i| = 4 \\ |y_{i+1} - y_i| = 3 \end{cases}{∣xi+1−xi∣=3∣yi+1−yi∣=4or{∣xi+1−xi∣=4∣yi+1−yi∣=3
Such a sequence is called a k-sequence.
If the first term of a k-sequence is (3,3)(3, 3)(3,3), find the possible second

Stable Reinforcement Learning with Icepop

Ring-1T achieves its exceptional performance through the inclusionAI team's Icepop algorithm and ASystem, a self-developed reinforcement-learning infrastructure purpose-built for trillion-parameter MoE models.

Icepop introduces masked bidirectional truncation correction, effectively narrowing the gap between training and inference and preventing the collapse seen in traditional GRPO algorithms.

This ensures long-term stability and consistent reasoning quality, even in extended training cycles.

Real-world Application Scenarios

Powered by trillion-scale reasoning and long-context understanding, Ring-1T helps developers, researchers, and enterprises tackle real-world tasks that demand accuracy, consistency, and deep analytical capability.

• Complex Problem Solving & Analytical Reasoning Use Ring-1T to handle advanced logic or quantitative tasks — from mathematical proof generation to structured analytical reports — with explainable reasoning chains and step-by-step clarity.

• Code Intelligence & System Optimization Integrate Ring-1T into your development workflows for multi-step code generation, debugging, and algorithm design. Its strong performance on LiveCodeBench and ICPC 2025 ensures reliable reasoning for production-grade coding agents.

• Agentic Workflows & Autonomous Planning Build capable AI agents that can plan, reason, and make decisions autonomously. Ring-1T's 131K context window and stable RL backbone enable consistent, multi-turn reasoning in complex real-world environments.

• Education & Knowledge-Driven Applications Power intelligent tutoring, training, or knowledge-assessment systems that explain logical and mathematical reasoning clearly — ideal for educational platforms and enterprise learning tools.

From advanced research to real-world deployment, Ring-1T brings deep reasoning and reliable intelligence to the next generation of applications — now fully available through SiliconFlow’s production-ready API.

Get Started Immediately

1. Explore: Try Ring-1T in the SiliconFlow playground.

2. Integrate: Use our OpenAI-compatible API. Explore the full API specifications in the SiliconFlow API documentation.

import requests

url = "https://api.siliconflow.com/v1/chat/completions"

payload = {
    "model": "inclusionAI/Ring-1T",
    "messages": [
        {
            "role": "user",
            "content": "Tell me a story"
        }
    ],
    "stream": True,
    "max_tokens": 4096,
    "enable_thinking": False,
    "thinking_budget": 4096,
    "min_p": 0.05,
    "stop": [],
    "temperature": 0.7,
    "top_p": 0.7,
    "top_k": 50,
    "frequency_penalty": 0.5,
    "n": 1,
    "response_format": { "type": "text" },
    "tools": [
        {
            "type": "function",
            "function": {
                "description": "<string>",
                "name": "<string>",
                "parameters": {},
                "strict": False
            }
        }
    ]
}
headers = {
    "Authorization": "Bearer <token>",
    "Content-Type": "application/json"
}

response = requests.post(url, json=payload, headers=headers)

print(response.json())

Ready to explore trillion-scale reasoning?

Access Ring-1T now on SiliconFlow and integrate deep, interpretable intelligence directly into your workflows.

Business or Sales Inquiries →

Join our Discord community now →

Follow us on X for the latest updates →

Explore all available models on SiliconFlow →