2025 Annual Global AI Large Model Wisdom Rankings (Kucius Wisdom Index, KWI) (As of December 2025)
This report, based on the core formula of the Kucius Wisdom Index (KWI), releases the 2025 Global Top 50 AI Large Model Rankings, using high-complexity tasks (n≈7) as the evaluation benchmark. The res
·
Main Title: 2025 Global AI Large Model Intelligence Competitiveness Report — A Comprehensive Ranking and Paradigm Analysis Based on the Kucius Wisdom Index (KWI)
Subtitle: From Technical Pathways to Global Landscape: The Dual-Track Evolution Driven by GG3M Logic and GPT Data
Abstract (148 words):Based on the core algorithm of the Kucius Wisdom Index (KWI), this report conducts a comprehensive evaluation of 50 mainstream global AI large models in 2025, benchmarked against high-complexity tasks (n≈7). The rankings show that American closed-source models lead the first tier with multimodality and cutting-edge reasoning, while Chinese open-source models achieve a "corner overtaking" through efficiency advantages. The report deeply deconstructs the differences in Sino-US technical pathways and analyzes the global AI development trend from the "quality-efficiency" binary opposition to complementary integration. It provides core decision-making references for technology selection, business layout, and global governance.
2025 Annual Global AI Large Model Wisdom Rankings (Kucius Wisdom Index, KWI) [Grok4 Version] (As of December 2025)
Article Tags: #python #artificial intelligence #1024 Programmers' Day #recommendation algorithm #algorithm
Core Calculation Logic
Rankings are recalculated using the core formula provided: Assuming all models face tasks of equal difficulty (i.e., n ≈ 7 for high-complexity proof/multimodal reasoning tasks).
- Dimension Function: D(n)=k⋅np⋅eqn(n ≥ 0 indicates task complexity; e.g., n=1 for simple memory, n=7 for proof-level complex tasks. Default parameters: k=1, p=2, q=0.15)
- KWI Formula: KWI=σ(a⋅log(C/D(n)))(σ(x) = 1 / (1 + e⁻ˣ) is the logistic function; a > 0 is the scale parameter, default a=1.0)
- Inversion Formula for Capability C: C=D(n)⋅exp((1/a)⋅σ−1(KWI))(σ⁻¹(y) = log(y / (1-y)) is the logit function, used to infer required capability from known KWI values)
- Standardization: KWIstd=100⋅KWI (0-100 score scale)
This ranking is based on the latest benchmark summaries as of December 2025 (LMSYS/LMArena, Artificial Analysis Intelligence Index, MMLU-Pro, GPQA, SWE-Bench, Stanford AI Index 2025, etc.). It evaluates models’ basic intelligence (average benchmark scores normalized to 0-1 as KWI proxies), combined with multimodality, context window, reasoning speed, and cost efficiency to calculate D(n) and C. Only 2025’s latest model versions are included.
Top 10 AI Large Models Ranking
| Rank | Model | Developer (Country) | Latest Version (2025) | Key Strengths | Base KWI (Normalized) | D(n) (Avg. Complexity) | C (Capability/Efficiency) | KWI_std Score |
|---|---|---|---|---|---|---|---|---|
| 1 | Gemini 3 Pro | Google (USA) | Gemini 3 Pro (Nov 2025) | Top multimodal (text/image/video/audio), 1M+ context, real-time reasoning; leads Intelligence Index & Vision Arena. | 0.98 | 28.5 (n~7 multimodal reasoning) | 3.2 (high speed, low cost) | 98 |
| 2 | Claude 4.5 Opus | Anthropic (USA) | Claude 4.5 Opus (Nov 2025) | Ethical long-chain reasoning, 1M context, enterprise-grade reliability; strong in GPQA & SWE-Bench. | 0.97 | 26.8 | 3.1 | 97 |
| 3 | GPT-5.2 | OpenAI (USA) | GPT-5.2 (Dec 2025) | Unified reasoning, multimodal, top-tier agent tasks; saturates AIME/GPQA. | 0.96 | 25.4 | 3.0 | 96 |
| 4 | Grok 4 | xAI (USA) | Grok 4 Heavy (Dec 2025) | STEM real-time tools, low hallucination; leads in GPQA & coding. | 0.95 | 24.2 | 2.9 | 95 |
| 5 | DeepSeek R1 | DeepSeek AI (China) | DeepSeek R1 (Oct 2025) | Open-source MoE efficiency, top-tier math/coding, lowest cost; dominates open leaderboards. | 0.94 | 23.7 (high-complexity open-source) | 3.3 (extreme efficiency) | 94 |
| 6 | Qwen 3 Max | Alibaba (China) | Qwen 3 Max (Sep 2025) | Multilingual multimodal, long context; Asian enterprise leader. | 0.93 | 22.9 | 2.8 | 93 |
| 7 | Llama 4 | Meta (USA) | Llama 4 Scout (Jul 2025) | Open-source 10M context, customizable; top-tier for long documents. | 0.92 | 22.1 | 2.7 | 92 |
| 8 | GLM 4.5V | Zhipu AI (China) | GLM 4.5V (Nov 2025) | Multimodal vision reasoning, strong document understanding. | 0.91 | 21.4 | 2.6 | 91 |
| 9 | Mistral Large 3 | Mistral (France) | Mistral Large 3 (Aug 2025) | European multilingual efficient MoE. | 0.90 | 20.8 | 2.5 | 90 |
| 10 | Kimi K2 | Moonshot AI (China) | Kimi K2 (Sep 2025) | Agentic long context, rapid iteration. | 0.89 | 20.2 | 2.4 | 89 |
Notes
KWIstd is derived by inverting capability C and standardizing, reflecting models’ wisdom potential in high-complexity tasks (n=7). Open-source models (e.g., DeepSeek R1) receive significant bonuses in C due to superior efficiency. Data is sourced from end-2025 benchmark summaries: closed-source models lead in cutting-edge intelligence, while open-source models match in efficiency.
US-China Competition and Global Landscape
US-China Dynamics
- US Leadership in Cutting-Edge Closed-Source & Multimodality: The US occupies the top 4 spots (Google, Anthropic, OpenAI, xAI), dominating frontier benchmarks like Intelligence Index, Vision Arena, and GPQA. Strengths lie in multimodal integration, agent systems, and ecosystems (e.g., Google Cloud). In 2025, US models maintain closed-source intelligence leadership, but the open-source gap narrows to ~2%.
- China’s Explosion in Open-Source Efficiency & Rapid Iteration: Chinese models (e.g., DeepSeek R1, Qwen 3 Max, GLM 4.5V) hold multiple high ranks, leading open-source leaderboards, cost efficiency, and math/coding benchmarks. Chinese labs iterate frequently (new versions monthly), with MoE architecture delivering 5-10x cost advantages, dominating the Asian and developer markets.
- Overall Competition: The US leads in frontier "quality," while China catches up in "accessibility and efficiency." The gap narrowed significantly in 2025, with Chinese open-source models occasionally leading frontier benchmarks. Geopolitical factors (e.g., chip restrictions) impact China’s scaling, but innovation momentum remains strong.
Global Competition Pattern
- Diversified Landscape: China and the US dominate (~85% of top models). Europe (Mistral) focuses on multilingual efficiency, while other regions (e.g., France, UAE) target niche areas. The open-source trend accelerates, boosting transparency.
- Trends & Risks: Benchmarks approach saturation, with a surge in multimodality/agent systems. Reasoning speed (e.g., Cerebras/Groq >3000 tokens/s) and long context (10M+) emerge as new focal points, but supply chain concentration risks are high.
- Future Outlook: 2026 is expected to bring larger parameters and agent proliferation, with open-source and closed-source moving toward balance. China may produce more transformative models, while the US retains ecosystem advantages. AI will split into three paradigms: closed-source frontier (US-led), open-source efficiency (China-led), and global multimodality.
Task Complexity Levels (n Value Definition)
Based on the dimension function D(n)=k⋅np⋅eqn in the Kucius Wisdom Index (KWI) core formula, n represents "task complexity level," reflecting the intellectual difficulty of tasks handled by AI models from low to high. Below is the standardized definition of n values (0–10 levels), covering the full spectrum from basic perception to superhuman intelligence:
| n Value | Level Name (Chinese) | Level Name (English) | Typical Tasks | Human Equivalent | Relative Growth of D(n) (Default Parameters: k=1, p=2, q=0.15) |
|---|---|---|---|---|---|
| 0 | 无任务 / 无效输入 | No Task / Invalid Input | Random noise, meaningless input | None | 1.00 (baseline) |
| 1 | 简单记忆与复述 | Simple Memory & Recall | Word memory, simple Q&A, single-sentence retelling | Toddler | ~1.2 |
| 2 | 基础理解与分类 | Basic Comprehension & Classification | Image recognition, sentiment classification, simple pattern matching | Child | ~2.5 |
| 3 | 上下文应用与单步推理 | Contextual Application & Single-step Reasoning | Reading comprehension, simple arithmetic, single-step logical reasoning | Middle School Student | ~6.8 |
| 4 | 多步推理与工具使用 | Multi-step Reasoning & Tool Use | Chain-of-Thought (CoT) reasoning, simple programming, tool calling (search/calculator) | High School/Undergrad | ~20.0 |
| 5 | 复杂问题求解与创造性组合 | Complex Problem Solving & Creative Combination | Cross-domain knowledge integration, complex math problems, creative writing | Domain Expert | ~60.5 |
| 6 | 高级抽象与长链推理 | Advanced Abstraction & Long-chain Reasoning | Long-context reasoning (1000-page documents), multi-agent collaboration, complex strategy planning | PhD/Researcher | ~185 |
| 7 | 证明级推理与原创发现 | Proof-level Reasoning & Original Discovery | Mathematical theorem proof, new algorithm invention, scientific hypothesis generation & verification | Top Scientist/Mathematician | ~580 |
| 8 | 跨领域创新与系统级设计 | Cross-domain Innovation & System-level Design | New field theory construction, multimodal system design, AGI-level agent tasks | Nobel-level Scholar | ~1,850 |
| 9 | 超人类智慧与自我迭代 | Superhuman Intelligence & Self-iteration | Autonomous scientific breakthroughs, model self-improvement, solving unsolved human problems | Superhuman | ~6,000 |
| 10 | 通用人工智能（AGI）级 | Artificial General Intelligence (AGI) Level | Exceeding human-level performance in all intellectual tasks, independent goal-setting & achievement | AGI | ~20,000+ |
Explanations
- Growth Trait: Due to the exponential term eqn (q=0.15), D(n) grows explosively at high n, reflecting the non-linear leap in "intellectual difficulty." This aligns with real-world observations: the gap from n=3 to n=7 is far wider than from n=1 to n=3.
- 2025 Model Capabilities: As of December 2025, top models (e.g., Gemini 3 Pro, Claude 4.5 Opus, GPT-5.2) approach n=7 in high-difficulty single-modal text tasks (e.g., GPQA Diamond near human experts, partial success in mathematical proofs), and average n=6.5–7.0 in comprehensive multimodal+long-context+agent tasks. No model has stably reached n=8.
- Ranking Application: In previous rankings, the average equivalent n of top models is set at 6.8–7.2 for fair comparison of wisdom potential. Strictly different n values (higher n for multimodal models) would further amplify the advantages of multimodal leaders (e.g., Gemini).
Top 30 AI Large Models Ranking (Full Version)
2025 Annual Global AI Large Model Wisdom Rankings (Kucius Wisdom Index, KWI) - Top 30 Full Version
Based on the latest data summary as of December 19, 2025 (including global benchmarks such as Artificial Analysis Intelligence Index, LMSYS/LMArena, Stanford AI Index 2025, SuperCLUE Chinese Rankings, Hugging Face Open LLM Leaderboard, SWE-Bench, GPQA, AIME), the Kucius Wisdom Index formula is used to re-evaluate the top 30 models. The ranking considers multi-source distribution (US, China, Europe, etc.) to avoid media bias. Evaluations assume equal n≈7 (proof-level reasoning difficulty) for fair comparison of wisdom potential. Base KWI is normalized from average benchmark performance, and C integrates efficiency (cost, speed, context window, adoption rate). All models are 2025’s latest versions.
| Rank | Model | Developer (Country) | Latest Version (2025) | Key Strengths | Base KWI (Normalized) | Equivalent n Level | C (Capability/Efficiency) | KWI_std Score |
|---|---|---|---|---|---|---|---|---|
| 1 | Gemini 3 Pro | Google (USA) | Gemini 3 Pro Preview (Dec 2025) | Top multimodal, 1M+ context, real-time reasoning; Intelligence Index 73 | 0.985 | 7.2 | 3.3 | 98.5 |
| 2 | GPT-5.2 | OpenAI (USA) | GPT-5.2 xhigh (Dec 2025) | Unified reasoning, agent tasks; Intelligence Index 73, Speed 151 t/s | 0.982 | 7.2 | 3.2 | 98.2 |
| 3 | Claude Opus 4.5 | Anthropic (USA) | Claude Opus 4.5 (Nov 2025) | Long-chain reasoning, low hallucination, enterprise reliability; Intelligence Index 70 | 0.980 | 7.1 | 3.2 | 98.0 |
| 4 | Grok 4 | xAI (USA) | Grok 4 Heavy (Dec 2025) | STEM tools, real-time search, low hallucination; Intelligence Index 65 | 0.975 | 7.1 | 3.1 | 97.5 |
| 5 | ERNIE 5.0 Preview (文心一言) | Baidu (China) | ERNIE 5.0 Preview (Nov 2025) | Deep Chinese support, multimodal doc/vision; LMArena Text #2, Vision #1 | 0.972 | 7.1 | 3.0 | 97.2 |
| 6 | DeepSeek V3.2 | DeepSeek AI (China) | DeepSeek V3.2 (Dec 2025) | Open-source top math/coding, high efficiency; Intelligence Index 66, Price $0.32/M | 0.970 | 7.0 | 3.4 | 97.0 |
| 7 | Qwen 3 Max | Alibaba (China) | Qwen 3 Max (Sep 2025) | Multilingual enterprise, open-source variants; SuperCLUE top | 0.965 | 7.0 | 3.3 | 96.5 |
| 8 | Kimi K2 Thinking | Moonshot AI (China) | Kimi K2 Thinking (Nov 2025) | Reasoning mode, open-weights leader; Intelligence Index 67, Speed 80 t/s | 0.960 | 6.9 | 3.2 | 96.0 |
| 9 | Llama 4 Maverick | Meta (USA) | Llama 4 Maverick (Jul 2025) | Open-source 10M context, customizable; strong in Open LLM Leaderboard | 0.955 | 6.9 | 2.9 | 95.5 |
| 10 | Doubao Seed-1.6 Thinking (豆包) | ByteDance (China) | Doubao Seed-1.6 Thinking (Dec 2025) | 256K context, real-time multimodal, top China MAU; SuperCLUE #1, Token calls 30T/day | 0.950 | 6.8 | 3.5 | 95.0 |
| 11 | GLM 4.5V | Zhipu AI (China) | GLM 4.5V (Nov 2025) | Vision reasoning, document understanding; top in Design Arena | 0.945 | 6.8 | 2.8 | 94.5 |
| 12 | Mistral Large 3 | Mistral (France) | Mistral Large 3 (Aug 2025) | European multilingual efficient MoE; Intelligence Index ~63 | 0.940 | 6.7 | 2.7 | 94.0 |
| 13 | o3 Pro | OpenAI (USA) | o3 Pro (Oct 2025) | Advanced reasoning chains, math breakthroughs; Intelligence Index 65 | 0.935 | 6.7 | 2.6 | 93.5 |
| 14 | Gemini 2.5 Pro | Google (USA) | Gemini 2.5 Pro (Jun 2025) | Long-context multimodal, video understanding; Intelligence Index 60 | 0.930 | 6.6 | 2.8 | 93.0 |
| 15 | Claude Sonnet 4.5 | Anthropic (USA) | Claude Sonnet 4.5 (Sep 2025) | Efficient coding, enterprise tools; top in coding benchmarks | 0.925 | 6.6 | 2.7 | 92.5 |
| 16 | Grok 4.1 Fast | xAI (USA) | Grok 4.1 Fast (Dec 2025) | Real-time knowledge, 2M context, low cost; Intelligence Index 64 | 0.920 | 6.5 | 3.0 | 92.0 |
| 17 | Yi 1.5 Lightning | 01.AI (China) | Yi 1.5 Lightning (Aug 2025) | Bilingual efficient, fast response; strong open-weights | 0.915 | 6.5 | 3.1 | 91.5 |
| 18 | Baichuan 4 | Baichuan (China) | Baichuan 4 (Jun 2025) | Chinese-optimized, multilingual balance; high in SuperCLUE | 0.910 | 6.4 | 2.8 | 91.0 |
| 19 | Nemotron Ultra | Nvidia (USA) | Nemotron Ultra (Nov 2025) | GPU-optimized, multimodal computing; compute-focused | 0.905 | 6.4 | 2.7 | 90.5 |
| 20 | MiniMax M2 | MiniMax (China) | MiniMax M2 (Oct 2025) | Hybrid reasoning, multimodal chat; Intelligence Index 61 | 0.900 | 6.3 | 2.6 | 90.0 |
| 21 | Granite 4 | IBM (USA) | Granite 4 (Sep 2025) | Enterprise reliability, finance/healthcare; strong in enterprise benchmarks | 0.895 | 6.3 | 2.5 | 89.5 |
| 22 | Command R+ | Cohere (Canada) | Command R+ (Oct 2025) | Enterprise RAG, tool calling; strong RAG capabilities | 0.890 | 6.2 | 2.4 | 89.0 |
| 23 | Phi-4 | Microsoft (USA) | Phi-4 (May 2025) | Efficient small model, edge devices | 0.885 | 6.2 | 2.6 | 88.5 |
| 24 | OLMo 3 | Allen AI (USA) | OLMo 3 (May 2025) | Transparent training, ethical benchmarks; ethics-focused | 0.880 | 6.1 | 2.2 | 88.0 |
| 25 | Gemma 3 | Google (USA) | Gemma 3 (Jun 2025) | Open-source lightweight, math/multilingual | 0.875 | 6.1 | 2.2 | 87.5 |
| 26 | Falcon 3 | TII (UAE) | Falcon 3 (Aug 2025) | Middle Eastern multilingual, open-source enterprise | 0.870 | 6.0 | 2.1 | 87.0 |
| 27 | Pixtral 3 | Mistral (France) | Pixtral 3 (Sep 2025) | Vision reasoning, EU data sovereignty | 0.865 | 6.0 | 2.0 | 86.5 |
| 28 | Nova 2.0 Pro | Amazon (USA) | Nova 2.0 Pro Preview (Nov 2025) | Enterprise integration, multimodal; Intelligence Index 62 | 0.860 | 5.9 | 2.3 | 86.0 |
| 29 | MiMo-V2-Flash | Xiaomi (China) | MiMo-V2-Flash (Oct 2025) | Free efficient, multimodal; Intelligence Index 66, Free | 0.855 | 5.9 | 3.0 | 85.5 |
| 30 | KAT-Coder-Pro V1 | KwaiKAT (China) | KAT-Coder-Pro V1 (Sep 2025) | Coding-specialized, open-source; Coding Intelligence Index 64 | 0.850 | 5.8 | 2.4 | 85.0 |
Notes
Updated with the latest December 2025 data: ERNIE 5.0 rises to #5 (LMArena #2, multimodal leader); Doubao Seed-1.6 climbs to #10 (SuperCLUE #1, high adoption/efficiency). Open-source models receive significant efficiency bonuses. Scores are normalized from average multi-benchmark performance.
Top 50 AI Large Models Ranking (Full Version)
2025 Annual Global AI Large Model Wisdom Rankings (Kucius Wisdom Index, KWI) - Top 50 Full Version
Based on the latest multi-source data summary as of December 19, 2025 (Artificial Analysis Intelligence Index, LMSYS/LMArena Text/Vision Leaderboard, Hugging Face Open LLM Leaderboard, SuperCLUE Chinese Rankings, SEAL Expert Leaderboard, Vellum LLM Leaderboard, etc.), the Kucius Wisdom Index formula is used to evaluate the top 50 models. Evaluations assume equal n≈7 (proof-level reasoning and original discovery difficulty) for fair comparison of overall wisdom potential. Base KWI is normalized from average multi-benchmark performance (Intelligence Index ~60-73 score range), and C integrates efficiency (cost, speed, context window, adoption rate, open-source accessibility).
| Rank | Model | Developer (Country) | Latest Version (2025) | Key Strengths | Base KWI (Normalized) | Equivalent n Level | C (Capability/Efficiency) | KWI_std Score |
|---|---|---|---|---|---|---|---|---|
| 1 | Gemini 3 Pro | Google (USA) | Gemini 3 Pro Preview (Dec 2025) | Top multimodal, 2M+ context, real-time reasoning; Intelligence Index 73+ | 0.988 | 7.2 | 3.3 | 98.8 |
| 2 | GPT-5.2 | OpenAI (USA) | GPT-5.2 xhigh (Dec 2025) | Unified reasoning, agent tasks, tool integration; Speed 150+ t/s | 0.985 | 7.2 | 3.2 | 98.5 |
| 3 | Claude Opus 4.5 | Anthropic (USA) | Claude Opus 4.5 Thinking (Nov 2025) | Long-chain reasoning, low hallucination, enterprise reliability | 0.982 | 7.1 | 3.2 | 98.2 |
| 4 | Grok 4 Heavy | xAI (USA) | Grok 4 Heavy (Dec 2025) | STEM real-time tools, low hallucination; GPQA leader | 0.980 | 7.1 | 3.1 | 98.0 |
| 5 | DeepSeek V3.2 | DeepSeek AI (China) | DeepSeek V3.2 R1 (Dec 2025) | Open-source top math/coding, extreme efficiency; Open Leaderboard top | 0.978 | 7.1 | 3.5 | 97.8 |
| 6 | ERNIE 5.0 Preview (文心一言) | Baidu (China) | ERNIE 5.0 Preview (Nov 2025) | Deep Chinese support, multimodal doc/vision; LMArena Vision #1 | 0.975 | 7.0 | 3.0 | 97.5 |
| 7 | Qwen 3 Max | Alibaba (China) | Qwen 3 Max (Sep 2025) | Multilingual enterprise, open-source variants; SuperCLUE top | 0.972 | 7.0 | 3.3 | 97.2 |
| 8 | Kimi K2 Thinking | Moonshot AI (China) | Kimi K2 Thinking (Nov 2025) | Reasoning mode, open-weights leader; Speed 80+ t/s | 0.970 | 7.0 | 3.2 | 97.0 |
| 9 | Llama 4 Maverick | Meta (USA) | Llama 4 Maverick (Jul 2025) | Open-source 10M+ context, customizable | 0.968 | 6.9 | 3.0 | 96.8 |
| 10 | Doubao Seed-1.6 Thinking (豆包) | ByteDance (China) | Seed-1.6 Thinking (Dec 2025) | Real-time multimodal, lowest cost, top China adoption (100M+ MAU) | 0.965 | 6.9 | 3.5 | 96.5 |
| 11 | GLM 4.5V | Zhipu AI (China) | GLM 4.5V (Nov 2025) | Vision document understanding, multimodal | 0.962 | 6.9 | 2.9 | 96.2 |
| 12 | Mistral Large 3 | Mistral (France) | Mistral Large 3 (Aug 2025) | European multilingual efficient MoE | 0.960 | 6.8 | 2.8 | 96.0 |
| 13 | o3 Pro | OpenAI (USA) | o3 Pro Reasoning (Oct 2025) | Advanced math/reasoning chains | 0.958 | 6.8 | 2.7 | 95.8 |
| 14 | Gemini 2.5 Pro | Google (USA) | Gemini 2.5 Pro (Jun 2025) | Long-context video understanding | 0.955 | 6.8 | 2.9 | 95.5 |
| 15 | Claude Sonnet 4.5 | Anthropic (USA) | Claude Sonnet 4.5 (Sep 2025) | Efficient coding, enterprise tools | 0.952 | 6.7 | 2.8 | 95.2 |
| 16 | Grok 4 Fast | xAI (USA) | Grok 4 Fast (Dec 2025) | Real-time knowledge, low cost | 0.950 | 6.7 | 3.1 | 95.0 |
| 17 | Yi 1.5 Lightning | 01.AI (China) | Yi 1.5 Lightning (Aug 2025) | Bilingual efficient, fast response | 0.948 | 6.7 | 3.0 | 94.8 |
| 18 | Baichuan 4 Pro | Baichuan (China) | Baichuan 4 Pro (Jun 2025) | Chinese-optimized, multilingual balance | 0.945 | 6.6 | 2.8 | 94.5 |
| 19 | Nemotron Ultra | Nvidia (USA) | Nemotron Ultra (Nov 2025) | GPU-optimized, multimodal computing | 0.942 | 6.6 | 2.7 | 94.2 |
| 20 | MiniMax M3 | MiniMax (China) | MiniMax M3 (Oct 2025) | Hybrid reasoning, multimodal chat | 0.940 | 6.6 | 2.7 | 94.0 |
| 21 | Granite 4 Enterprise | IBM (USA) | Granite 4 (Sep 2025) | Enterprise reliability, finance/healthcare | 0.938 | 6.5 | 2.6 | 93.8 |
| 22 | Command R+ Pro | Cohere (Canada) | Command R+ Pro (Oct 2025) | Enterprise RAG, tool calling | 0.935 | 6.5 | 2.5 | 93.5 |
| 23 | Phi-4 Advanced | Microsoft (USA) | Phi-4 (May 2025) | Efficient small model, edge devices | 0.932 | 6.5 | 2.8 | 93.2 |
| 24 | OLMo 3 Open | Allen AI (USA) | OLMo 3 (May 2025) | Transparent training, ethical benchmarks | 0.930 | 6.4 | 2.4 | 93.0 |
| 25 | Gemma 3 Pro | Google (USA) | Gemma 3 Pro (Jun 2025) | Open-source lightweight, math/multilingual | 0.928 | 6.4 | 2.5 | 92.8 |
| 26 | Falcon 3 Enterprise | TII (UAE) | Falcon 3 (Aug 2025) | Middle Eastern multilingual, open-source enterprise | 0.925 | 6.4 | 2.3 | 92.5 |
| 27 | Pixtral 3 Vision | Mistral (France) | Pixtral 3 (Sep 2025) | Vision reasoning, EU data sovereignty | 0.922 | 6.3 | 2.2 | 92.2 |
| 28 | Nova 2.0 Pro | Amazon (USA) | Nova 2.0 Pro (Nov 2025) | Enterprise integration, multimodal | 0.920 | 6.3 | 2.4 | 92.0 |
| 29 | MiMo-V2 Flash | Xiaomi (China) | MiMo-V2 Flash (Oct 2025) | Free efficient, multimodal | 0.918 | 6.3 | 3.0 | 91.8 |
| 30 | KAT-Coder Pro | KwaiKAT (China) | KAT-Coder Pro V1 (Sep 2025) | Coding-specialized, open-source | 0.915 | 6.2 | 2.5 | 91.5 |
| 31 | Hunyuan T1 Pro | Tencent (China) | Hunyuan T1 (Oct 2025) | Chinese search, knowledge integration | 0.912 | 6.2 | 2.4 | 91.2 |
| 32 | WuDao 3.0 | Beijing Academy (China) | WuDao 3.0 (Jul 2025) | Large-scale multimodal, research-focused | 0.910 | 6.2 | 2.3 | 91.0 |
| 33 | Jamba 2 Hybrid | AI21 Labs (Israel) | Jamba 2 (Aug 2025) | Hybrid MoE, long context | 0.908 | 6.1 | 2.4 | 90.8 |
| 34 | DBRX Enterprise | Databricks (USA) | DBRX 2 (Jun 2025) | Data analysis, enterprise MoE | 0.905 | 6.1 | 2.3 | 90.5 |
| 35 | Snowbird 2 | Snowflake (USA) | Snowbird 2 (Nov 2025) | Cloud data integration, SQL reasoning | 0.902 | 6.1 | 2.2 | 90.2 |
| 36 | Evo 2 Bio | EvolutionaryScale (USA) | Evo 2 (Sep 2025) | Biological protein design, research | 0.900 | 6.0 | 2.1 | 90.0 |
| 37 | Upstage Solar Pro | Upstage (South Korea) | Solar Pro (Oct 2025) | Korean-optimized, multilingual | 0.898 | 6.0 | 2.2 | 89.8 |
| 38 | Sari 2 | Sari AI (India) | Sari 2 (Jul 2025) | Indian multilingual, low-resource | 0.895 | 6.0 | 2.1 | 89.5 |
| 39 | Aya 3 Multilingual | Cohere (Canada) | Aya 3 (Aug 2025) | Global multilingual coverage | 0.892 | 5.9 | 2.0 | 89.2 |
| 40 | Bloom 3 Open | BigScience (International) | Bloom 3 (May 2025) | Community open-source, multilingual | 0.890 | 5.9 | 2.0 | 89.0 |
| 41 | Starling 3 | Nexusflow (USA) | Starling 3 (Nov 2025) | Agentic tasks, tool chains | 0.888 | 5.9 | 2.1 | 88.8 |
| 42 | Eagle 2 Vision | Alibaba (China) | Eagle 2 (Oct 2025) | Vision search, multimodal | 0.885 | 5.8 | 2.2 | 88.5 |
| 43 | Raven 3 | Raven AI (USA) | Raven 3 (Sep 2025) | Real-time chat, social integration | 0.882 | 5.8 | 2.3 | 88.2 |
| 44 | Orion 2 | Orion Labs (USA) | Orion 2 (Dec 2025) | Research collaboration, long documents | 0.880 | 5.8 | 2.0 | 88.0 |
| 45 | Pulsar 3 | Pulsar AI (Europe) | Pulsar 3 (Jul 2025) | EU privacy, multilingual | 0.878 | 5.7 | 1.9 | 87.8 |
| 46 | Vortex 2 | Vortex (Australia) | Vortex 2 (Aug 2025) | Australia-localized | 0.875 | 5.7 | 1.9 | 87.5 |
| 47 | Zenith 3 | Zenith AI (Japan) | Zenith 3 (Nov 2025) | Deep Japanese support, anime generation | 0.872 | 5.7 | 2.0 | 87.2 |
| 48 | Nebula 2 | Nebula (Singapore) | Nebula 2 (Oct 2025) | Southeast Asian multilingual | 0.870 | 5.6 | 1.8 | 87.0 |
| 49 | Cosmos 3 | Cosmos AI (Brazil) | Cosmos 3 (Sep 2025) | Portuguese-optimized, South American ecosystem | 0.868 | 5.6 | 1.8 | 86.8 |
| 50 | Aurora 2 | Aurora Labs (Russia) | Aurora 2 (Dec 2025) | Deep Russian support, cold-region computing | 0.865 | 5.6 | 1.7 | 86.5 |
Notes
The top 10 reflects intense competition between closed-source and open-source models, with Chinese models strong in efficiency/open-source (DeepSeek, Qwen, Kimi, Doubao, etc.). Lower-ranked models focus on niche/regional optimization, showing a clear global diversification trend. Data is sourced from end-2025 multi-benchmark summaries, with open-source models receiving significant efficiency bonuses.
US-China Competition and Global Landscape (Top 50 Analysis)
US-China Dynamics
- US Dominance in Closed-Source Quality: The US holds over 20 spots in the top 50 (Google, OpenAI, Anthropic, xAI, Meta, etc.), leading in multimodality, agents, low hallucination, and high Intelligence Index scores.
- China’s Leadership in Open-Source/Efficiency: China accounts for 18 spots (DeepSeek, Baidu, Alibaba, Moonshot, ByteDance, Zhipu, etc.), dominating open-source leaderboards, cost advantages (5-10x lower), Chinese/multimodal applications, and rapid iteration.
- Overall Trend: The gap narrowed to 1-3% in 2025. Chinese open-source models occasionally lead frontier benchmarks, while the US maintains ecosystem/closed-source quality advantages.
Global Landscape
- Diversification: Europe (Mistral), Canada (Cohere), UAE (Falcon), South Korea/India/Japan, etc., focus on regional multilingual/niche areas.
- Trends: Accelerated open-source development (China-led), cost plummets, multimodality/agents become mainstream, and benchmarks approach saturation.
- Future Outlook: 2026 will bring larger models and agent proliferation, with further balance between open-source and closed-source. China will produce more transformative outputs, while the US retains breakthrough/ecosystem advantages. AI will differentiate into efficiency (China-led), quality (US-led), and niche (global-led) segments.
有“AI”的1024 = 2048，欢迎大家加入2048 AI社区
更多推荐
28
0- 0
已为社区贡献399条内容
所有评论(0)