Medium

Technology

ChatGPT

Large Language Models

Machine Learning

Tech Companies

3 min readOct 13, 2025

Today, I saw the Ant Ling team released Ring-1T, an open-source hundred-million parameter thinking model based on the Ling 2.0 architecture.

The model’s parameter size reaches a total of 1 trillion, with 50 billion active parameters and a context window of 128K. It employs Icepop RL and ASystem for reinforcement learning training, achieving the best performance among open-source models in natural language reasoning.

This is a chart showing the performance of different AI models in multiple benchmark tests. The chart includes eight different test items: AIME 25, CodeForces, ARC-AGI-1, Arena-Hard-v2.0, HMMT 25, LiveCodeBench(2408-2505), HealthBench, and Creative Writing v3. Each test item has a specific metric to measure the model's performance, such as pass@1, rating, win-rate, and score. These metrics are displayed in the form of bar charts, where blue represents the performance of the Ring-1T Open-Weights model, light gray represents the performance of the Ring-1T-Preview Open-Weights model, dark gray represents the performance of the Gemini-2.5-Pro Close-Weights model, purple represents the performance of the DeepSeek-V3.1-Termius-Thinking Open-Weights model, green represents the performance of the Qwen3-235B-A22B-Thinking-2507 Open-Weights model, and orange represents the performance of the GPT5-Thinking (High) Close-Weights model. From the chart, it can be seen that the Ring-1T Open-Weights model has achieved high scores in most test items, especially standing out in the AIME 25 and CodeForces tests. Other models have shown certain competitiveness in some test items, but overall, they are not as excellent as the Ring-1T Open-Weights model.

What is particularly impressive is its performance in the 2025 IMO test — solving P1, P3, P4, and P5 problems in one go, and the P2 problem was also solved after three attempts, achieving a silver medal level through pure language reasoning. The team mentions they are still continuing training, with the goal of reaching gold level.

In terms of technical details, ASystem, as a high-performance RL framework, supports training at the trillion level, featuring unified memory pools, transparent offloading, and direct GPU point-to-point communication. Its AReaL framework supports server sandbox environments for 10+ languages, with a throughput of 10K/s.

This is a line chart showing the trend of two different datasets over 400 steps. The "GRPO" dataset is represented in purple, while the "+IcePop" dataset is in green. The x-axis represents steps (Step), ranging from 0 to 400, and the y-axis represents the absolute value of the log probability difference (abs(log p_infer - log p_train)), ranging approximately from 0.02650 to 0.02850. The chart shows that the GRPO data curve has significant fluctuations, especially a noticeable downward trend in the early stages, gradually stabilizing but still with minor fluctuations. In contrast, the +IcePop data curve is relatively smooth with smaller fluctuations, indicating a more stable trend. There is a legend box in the upper right corner of the chart, indicating the sources of the two datasets.

This is a line chart showing the performance of the GRPO and +IcePop methods during training. The x-axis represents steps (Step), ranging from 0 to 400, and the y-axis represents the difference between the maximum log likelihood function value and the log likelihood function value (max(log p_infer - log p_train)), ranging from 10 to 26. The purple line represents the GRPO method, and the green line represents the +IcePop method. From the chart, it can be seen that the GRPO method's curve fluctuates significantly throughout the process, while the +IcePop method's curve is relatively stable. At around 400 steps, the GRPO method's value suddenly rises sharply, exceeding the +IcePop method. The chart has a legend in the upper left corner explaining what different colored lines represent.

The model is already available on HuggingFace, including the FP8 version. Florian S mentioned that this open-source model outperforms Sonnet 4.5 by 7 points on LiveCodeBench, and Jon Durbin also noted that the model has been launched on the Chutes platform.

This is a screenshot of a scoring report for a math problem, divided into five problems, each with a correct answer, the Ring-1T system's response, and corresponding evaluations. 1. The correct answer to the first problem is k=0,1,3, and the Ring-1T preview (three attempts) gives the answer that k can be 0 or odd integers between 1 and n (partially correct), with a single attempt giving k=0,1,3 (correct). 2. The second problem is scored based on the standard answer and the proof process provided by Ring-1T, scoring 90/100. The comment notes that this is a completely correct solution in terms of logical path and mathematical accuracy, with the only slight drawback being the lack of a full demonstration of the final algebraic simplification steps during the presentation. 3. The correct answer to the third problem is 4, and the Ring-1T preview (single attempt) also gave 4 (correct), with the single attempt also giving 4 (correct). 4. The correct answer to the fourth problem is $2^a3^bM$, where a is a positive odd number, b is greater than or equal to (a+1)/2, and all prime factors are greater than or equal to 7. The Ring-1T preview (three attempts) has the correct expression form but incorrect conditions, while the single attempt gives the expression 6×12^n×k (n>=0 and k is an odd number not divisible by 5) (correct, equivalent expression). 5. The correct answer to the fifth problem is the critical value of λ being $\frac{1}{\sqrt{2}}$, with Alice winning if λ is greater than this value, Bazza winning if λ is smaller, and a tie if equal. The Ring-1T preview (three attempts) has incorrect conditions, while the single attempt gives conditions consistent with the correct answer (correct).

This is an image with a gradient background, transitioning from teal at the top to blue, then to orange. In the upper left corner, there is a small icon showing a white human figure pattern on a black background. Next to it, the text "inclusionAI" is displayed. In the center, large font reads "/Ring-1T", below which is a yellow smiley emoji and a link address "huggingface.co".

This is an image with a gradient background, transitioning from teal, blue, purple, and orange from top to bottom. In the upper left corner, there is a small icon showing a black rectangle with a human figure and a white triangle. To the right of the icon, there is white text "inclusionAI". Below, there is a large white text line reading "/Ring-1T-FP8", with a yellow smiley emoji in the lower right corner, next to a white website address "huggingface.co".

Interestingly, some users mentioned that O3 mini has already reached 2000 points on CodeForces, O3 full is close to 3000 points, while GPT-5 High only has 1918 points. This makes one reflect on the differences in performance between different models on specific tasks.

The model is currently accessible via ZenMux Chat/API. The team mentions this is their first trillion-level deep thinking model, still under continuous training and evolution. For those interested in the development of open-source large models, this is indeed a noteworthy advancement.

Link:Source

Written by AI Engineering