Trillion-Scale Reasoning Model: Ant Group’s First Open Source Release with 20 Trillion Tokens Disrupts Open AI
New Intelligence Report: Ant Group Launches Ring‑1T
Editor's Note:
Ant Group has unveiled the trillion-parameter reasoning model Ring‑1T, setting new open-source SOTA records in math competitions, logical reasoning, and medical Q&A. Tests show that Ring‑1T's reasoning approaches closed-source leaders — heralding the trillion-parameter era for open-source AI.
---
Ring‑1T Achieves Breakthrough Performance
On October 14, Ant Group officially released Ring‑1T, a trillion-parameter reasoning model, with impressive results in:
- Mathematics Competitions: AIME 25, HMMT 25
- Code Generation: CodeForces
- Logical Reasoning: ARC‑AGI‑v1
- Medical Q&A: HealthBench benchmark
Key Highlights
- On Arena‑Hard‑v2 and CreativeWriting‑v3, Ring‑1T ranks among the very best open-source reasoning models alongside DeepSeek and Qwen.
- 81.59% success rate on Arena-Hard V2 — almost matching the closed-source GPT‑5‑Thinking(High) score of 82.91%.
- Demonstrates balanced capability across reasoning-intensive and creative benchmarks.
> Bottom line: Open-source AI is now capable of competing directly with closed-source giants.
---
Try Ring‑1T Yourself
Access Ring‑1T via Ant’s Treasure Box:
Model Download Links:
- HuggingFace: https://huggingface.co/inclusionAI/Ring-1T
- ModelScope: https://modelscope.cn/models/inclusionAI/Ring-1T
---
Ant’s Push Toward AGI
September Releases
Ant Group launched seven models in September:
- Ring‑1T‑preview
- Ring‑flash‑linear‑2.0
- Ring‑flash‑2.0
- Ling‑flash‑2.0
- Ming‑lite‑omni‑1.5
- Ring‑mini‑2.0
- Ling‑mini‑2.0
Two Trillion‑Parameter Models in October
- Oct 9: Ling‑1T (general-purpose trillion-parameter model)
- Oct 14: Ring‑1T (reasoning-focused trillion-parameter model)
User Feedback:
Ling‑1T has been praised online as outperforming DeepSeek, Gemini, and o3-mini.
---
Technical Foundations
Model Architecture & Training
- Ring‑1T shares the Ling‑1T architecture but with:
- 20T high-quality corpora
- Reinforcement learning tuned for reasoning skills
- Enhanced decontamination filtering to avoid training data leakage
- Benchmark improvements over preview:
- Arena-hard-v2: +8.18%
- ARC-AGI-v1: +5.14%
- HealthBench: +3.49%
---
Benchmark Achievements
International Mathematical Olympiad (IMO 2025)
- Integrated into AWorld multi-agent framework
- Achieved silver medal level: solved Problems 1, 3, 4, 5 on first attempt
- GitHub project: https://github.com/inclusionAI/AWorld
ICPC 2025 World Finals
- Outperformed Gemini 2.5 Pro in programming tasks
- Reasoning traces open-sourced:
- https://github.com/inclusionAI/AWorld/tree/main/examples/imo/samples/samples%20from%20Ring-1T
---
Hands-On Testing
Simulation Tasks
- Earth–Mars Flight: 3D three.js simulation with parameter controls
- Physics Simulation: "Neon Collider" — complex collision physics with HTML5 Canvas
- Space Invaders Game: Comparable visuals to Gemini 3, superior to Gemini 2.5
- Cryptarithmetic Puzzle: Solved BASE + BALL = GAMES with systematic enumeration
- Math Problems: Solved indefinite integrals involving symbolic manipulation
---
Creative Writing Experiments
- Generated poetry-style prose themed on AGI
- Wrote a Mount Everest piece in classic Chinese literary style
---
Engineering Innovations
Ling 2.0 Optimizations
- High-sparsity MoE with 1/32 expert activation
- FP8 mixed precision
- MTP acceleration
- 20T tokens of high-quality data, with >40% reasoning-related content
Training Innovations
- IcePop Algorithm — stabilizes RL training in MoE models with double-sided clipping & masking
- ASystem RL Platform — scales RL training from 10B to 1T parameters
- AReaL Framework — open-source fully asynchronous RL training for agents
- GitHub: https://github.com/inclusionAI/AReaL
---
Community & Ecosystem Impact
Platforms like AiToEarn官网 integrate:
- AI content generation
- Multi-platform publishing (Douyin, Kwai, Bilibili, Facebook, Instagram, YouTube, X, etc.)
- Analytics & monetization tools
- AI model ranking: https://rank.aitoearn.ai
---
Conclusion
Ring‑1T signals:
- A new tier for open-source AI reasoning
- Cooperative synergy between large-scale models and creator platforms
- Technical stability through innovations like IcePop & ASystem
AGI is approaching — and with Ring‑1T, open-source is ready.
---
References:
- https://x.com/AntLingAGI
- https://ringtech.notion.site/icepop
- https://ringtech.notion.site/Small-Leak-Can-Sink-a-Great-Ship-Boost-RL-Training-on-MoE
---
If you’d like, I can also create a compact, executive summary version of this report so decision-makers can digest the key stats, benchmarks, and ecosystem impact in under 2 minutes.
Do you want me to prepare that?
