⚛️量子位•較早收集於 74m
懂人性更懂執行，螞蟻這個萬億開源模型把情商和Agent戰鬥力都給拉滿了
💡Free trillion-param open model dominates agents & EQ - game-changer for builders
⚡ 30-Second TL;DR
有什麼變化
萬億參數開源LLM發布
為什麼重要
開源萬億規模AI，賦能Agent應用，挑戰封閉巨頭。
下一步行動
Download Ant's trillion-param model from Hugging Face and test agent benchmarks.
誰應關注：Developers & AI Engineers
關鍵要點
- •萬億參數開源LLM發布
- •先進類人情商與理解力
- •卓越Agent執行能力
- •巨型規模下高效運作
🧠 深度解析
背景與延伸：來自公開資料，非原文內容。引用 6 個來源。
🔑 增強重點摘要
- •Ant Group open-sourced Ring-2.5-1T, the world's first trillion-parameter reasoning model using a hybrid linear architecture, excelling in long-text generation, mathematical reasoning, and agent task execution.[1][2]
- •Ring-2.5-1T achieves leading open-source performance in benchmarks like IMOAnswerBench, HMMT-25, LiveCodeBench-v6, IMO 2025 (35/42, gold standard), and CMO 2025 (105/126).[1][3][4]
- •The model demonstrates superior efficiency, with throughput advantages over KIMI K2 (32B active params) in long-sequence tasks, reducing memory access by over 10x and boosting throughput 3x+ for lengths >32K.[1][2]
- •Ring-2.5-1T is part of Ant Group's BaiLing (Ling) family evolution, alongside Ling-2.5-1T (1M token context, efficient reasoning) and multimodal Ming-Flash-Omni-2.0, available on Hugging Face and ModelScope.[3][4]
- •Improvements over prior Ring-1T include better generation efficiency, cognitive depth, and long-range execution, supporting AGI efforts.[2][3][4]
📊 競品分析▸ Show
| Feature | Ant Ring-2.5-1T | KIMI K2 | Qwen 3.5 |
|---|---|---|---|
| Parameters | 1T (hybrid linear) | 1T (32B active) | Not specified |
| Strengths | Math reasoning (IMO 35/42), agents, long-text efficiency | Coding, visual | Reasoning, coding, agents (multimodal) |
| Efficiency | 3x+ throughput >32K, 10x less memory | Lower throughput in long seq | Not detailed |
| Benchmarks | Gold on IMO/CMO 2025, LiveCodeBench-v6 | Not directly compared | Not directly compared |
| Pricing | Open-source (free) | Subscription up to $1,908/yr | Not detailed |
🛠️ 技術深入
- Hybrid linear architecture enables efficient long-sequence reasoning, outperforming traditional models in throughput as generation length increases.[1][2]
- Achieves IMO 2025: 35/42 (gold medal), CMO 2025: 105/126 (surpasses national cutoff), AIME 2026 efficiency with ~5,890 tokens vs. 15k-23k for frontiers (related Ling-2.5).[3][4]
- Heavy Thinking mode excels in math competitions (IMOAnswerBench, HMMT-25), code gen (LiveCodeBench-v6), logical reasoning, agent tasks.[1]
- Supports 1M token context (Ling-2.5 counterpart), native agent interaction, fine-grained preference alignment.[3][4]
- Open-sourced on Hugging Face/ModelScope under open licenses.
🔮 前景展望AI analysis grounded in cited sources
Ant Group's trillion-parameter open-source models like Ring-2.5-1T advance agentic AI and reasoning efficiency, providing high-performance foundations for complex tasks and intensifying competition in China's AI ecosystem toward AGI, while enabling broader industry adoption via efficiency gains over closed alternatives.[1][3][4]
⏳ 時間線
2025-10
Ling 2.0 series unveiled, foundation for Ling-2.5 and Ring-2.5 evolutions.[3][4]
2026-02
Ring-1T released, precursor with improvements leading to Ring-2.5-1T.[2]
2026-02-11
Ming-Flash-Omni-2.0 released, unifying speech/audio/music in BaiLing family.[3][4]
2026-02-13
Ring-2.5-1T open-sourced, world's first hybrid linear trillion-param reasoning model.[1]
📎 來源 (6)
Factual claims are grounded in the sources below. Forward-looking analysis is AI-generated interpretation.
- news.aibase.com — 25520
- aastocks.com — Latest News
- afp.com — Ant Group Releases Ling 25 1t and Ring 25 1t Evolving Its Open Source AI Model Family
- businesswire.com — Ant Group Releases Ling 2.5 1t and Ring 2.5 1t Evolving Its Open Source AI Model Family
- chinatalk.media — Chinese AI Rings in the Year of the
- fintechweekly.com — Agile Infrastructure AI Insurance Aspida
📰
AI 週報
閱讀本週精選 AI 大事摘要 →
👉相關動態
AI 策展新聞聚合。所有內容版權歸原始發布者所有。
原始來源： 量子位 ↗
每週 AI 簡報
每週一封，可隨時退訂。