🔥36氪•較早收集於 42m
智譜尾盤漲幅一度擴大至37%，總市值突破3000億港元
💡Zhipu AI hits $38B cap surge—track GLM model updates & China AI rivalry
⚡ 30-Second TL;DR
有什麼變化
尾盤漲幅達37%
為什麼重要
提升智譜GLM模型能見度，於中國AI競爭中。可能迫使競爭對手調整定價，並預示研發擴張資金。AI創辦人應留意合作估值變化。
下一步行動
Benchmark Zhipu GLM-4o API costs against GPT-4o post-surge funding signals.
誰應關注：Founders & Product Leaders
關鍵要點
- •尾盤漲幅達37%
- •總市值超3000億港元
- •發生於2月20日
- •凸顯中國AI股票動能
🧠 深度解析
背景與延伸：來自公開資料，非原文內容。引用 5 個來源。
🔑 增強重點摘要
- •Zhipu AI's shares surged up to 37% on February 20, 2026, pushing market cap beyond 300 billion HKD, following the February 11 launch of GLM-5 which drove earlier gains of 29-42%[2][3]
- •GLM-5, a 744-745 billion parameter MoE model with 44 billion active parameters, ranks 4th globally, outperforming Gemini 3 Pro and Kimi K2.5, matching Claude Opus 4.5 in coding[1][2][5]
- •Trained entirely on Huawei Ascend chips using MindSpore, GLM-5 achieves full hardware independence, a milestone for China's AI self-reliance[2][5]
- •Zhipu raised ~4.35B HKD (558M USD) in its January 8, 2026 HK IPO, despite H1 2025 net loss of 2.36B yuan with R&D doubling to 1.6B yuan[1][5]
- •GLM-5 integrates DeepSeek Sparse Attention and asynchronous agent RL for enhanced agentic tasks, coding, and integration with OpenClaw[3]
📊 競品分析▸ Show
| Feature/Benchmark | Zhipu GLM-5 | Anthropic Claude Opus 4.5 | Google Gemini 3 Pro | Moonshot Kimi K2.5 |
|---|---|---|---|---|
| Global Rank | 4th | Top 3 | Outperformed | Outperformed |
| Parameters | 744-745B total, 44B active | N/A | N/A | N/A |
| Coding/Programming | Matches Opus 4.5 | Strong benchmark | N/A | Lower than GLM-5 |
| Training Hardware | Huawei Ascend only | N/A | N/A | N/A |
| Availability in China | Yes, raised coding plan 30% | Unavailable | N/A | Competitor |
🛠️ 技術深入
- Mixture-of-Experts (MoE) architecture: 745B total parameters, 44B active per inference, 256 experts with 8 activated per token[2][5]
- Integrates DeepSeek Sparse Attention for efficient long-text processing without quality loss, reducing deployment costs[3]
- Asynchronous agent reinforcement learning for continuous learning from long-range interactions[3]
- Native Agent Mode for autonomous task breakdown, multi-step reasoning, coding, creative writing, and generating office documents (.docx, .pdf, .xlsx)[4]
- Trained exclusively on Huawei Ascend chips with MindSpore framework, open-source under MIT license[2][5]
- Supports domestic Chinese chips, optimized for complex systems engineering and OpenClaw integration[1][3]
🔮 前景展望AI analysis grounded in cited sources
Zhipu AI's stock surge and GLM-5 launch signal accelerating momentum in Chinese AI, demonstrating viable alternatives to US hardware dominance and positioning firms like Zhipu as global contenders in open-source LLMs with agentic capabilities[1][2][4][5].
⏳ 時間線
2019-01
Zhipu AI founded as Tsinghua University spinoff
2025-01
H1 2025: Reported 2.36B yuan net loss, R&D expenses double to 1.6B yuan
2025-09
DeepSeek launches Sparse Attention mechanism later integrated into GLM-5
2025-11
Anthropic launches Claude Opus 4.5
2026-01
HK IPO on Jan 8 raises 4.35B HKD to fund AI development
2026-02
GLM-5 launched Feb 11, shares surge 29-42%, later 37% to 300B HKD market cap on Feb 20
📎 來源 (5)
Factual claims are grounded in the sources below. Forward-looking analysis is AI-generated interpretation.
- caixinglobal.com — Zhipu AI Launches New Model Better at Coding Learning 102413996
- siliconrepublic.com — Zhipu Glm 5 Chinese AI Start Up Artificial Intelligence
- yicaiglobal.com — Zhipu AI Pops After Chinese Startup Launches New Foundation Model
- evrimagaci.org — Zhipu AI Unveils Glm5 Model Redefining Global AI Race 528618
- trendingtopics.eu — Glm 5 the Worlds Strongest Open Source LLM Solely Trained on Chinese Huawei Chips
📰
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： 36氪 ↗
每週 AI 簡報
每週一封，可隨時退訂。