大模型全景图
104一站式了解所有主流大模型，覆盖 OpenAI、Anthropic、Google、DeepSeek 等提供商。
Model
Specs
ChatGLM-4-9B
Open Source
文本 智谱 AI
ChatGLM-4-9B is a lightweight open-source model launched by Zhipu AI, achieving excellent comprehensive capabilities at 9B parameters. It supports 128K context and multilingual dialogue, and excels in Agent tool calling, code generation, and mathematical reasoning. ChatGLM-4-9B can run efficiently on consumer-grade GPUs, making it a popular choice for developers in AI application development.
Gemini 3.8 Flash
闭源
多模态 Google
Gemini 3.8 Flash, released by Google in September 2026, is the newest generation of the Flash family, launched alongside Gemini 3.8 Flash Cyber for cybersecurity scenarios, continuing the 3.x line's coding and agentic capabilities.
o3
Commercial Closed-source
推理 OpenAI
o3 is OpenAI's most powerful reasoning model, employing chain-of-thought deep reasoning mechanisms, achieving unprecedented levels in mathematics, programming, and scientific reasoning. It achieved excellent results in the 2025 IOI gold-medal-level programming competition and surpassed human experts on the GPQA doctoral-level scientific reasoning benchmark. Suitable for complex reasoning tasks requiring deep thought.
Qwen3
Open Source
文本 阿里云
Qwen3 is the latest open-source model series released by Alibaba's Tongyi Qianwen team, offering specifications from 0.6B to 235B. The flagship Qwen3-235B uses MoE architecture, benchmarking against DeepSeek-V3 in coding, mathematics, and general capabilities. The Qwen3 series particularly strengthens Thinking Mode, supports tool calling and Agent capabilities, and is one of the most competitive model series in the open-source community.
Claude 4 Sonnet
Commercial Closed-source
文本 Anthropic
Claude 4 Sonnet is Anthropic's mid-to-high-end model, striking the best balance between intelligence and efficiency. It excels in programming, writing, and complex analysis tasks, supporting a 200K token context window. Claude 4 Sonnet achieves new breakthroughs on the SWE-bench coding benchmark while maintaining Anthropic's consistently high standards in writing quality and safety.
Muse Glimmer
Open Source
多模态 Meta
Muse Glimmer, released open-source by Meta's Superintelligence Labs on August 10, 2026, marks Meta's return to the open-weights track, publishing model weights openly as part of its Personal Superintelligence foundation-model strategy.
Gemini 2.5 Pro
Commercial Closed-source
多模态 Google
Gemini 2.5 Pro is Google's most advanced thinking model, capable of deep reasoning before answering. It has an ultra-long context window of 1 million tokens and natively supports multimodal understanding of text, images, video, audio, and code. It has long been at the top of the LMArena leaderboard, with particularly outstanding mathematical and scientific reasoning capabilities, suitable for professional scenarios requiring complex reasoning and cross-modal understanding.
Claude 3.5 Haiku
Commercial Closed-source
文本 Anthropic
Claude 3.5 Haiku is Anthropic's fastest model, designed for scenarios requiring low latency and high throughput. While maintaining high reasoning quality, it responds extremely quickly, making it ideal for real-time conversations, content moderation, and data processing. Among models of the same speed class, Claude 3.5 Haiku leads in intelligence.
GPT-5.5 Instant
Commercial Closed-source
文本 OpenAI
GPT-5.5 Instant is a flagship free model launched by OpenAI in May 2026, accessible to everyone with zero barriers. It maintains the core capabilities of GPT-5.5 while achieving extremely low inference latency, covering all scenarios such as writing, programming, and analysis. It is OpenAI's most inclusive flagship model ever.
Yi-Lightning
Commercial Closed-source
文本 零一万物
Yi-Lightning is a high-speed large model launched by 01.AI, adopting a MoE architecture, achieving extremely low inference latency while maintaining high-quality output. It is known for fast response on the LMArena leaderboard and performs particularly well in Chinese scenarios. Yi-Lightning is positioned as cost-effective, suitable for real-time dialogue and customer service scenarios requiring quick responses.
GPT-6 Sol
Commercial Closed-source
推理 OpenAI
GPT-6 Sol is the mid-tier model of OpenAI's GPT-6 family, released on September 22, 2026 alongside GPT-6 Luna, with noticeably stronger reasoning than Luna. It accepts text and image input, offers roughly a 1.05M-token context and 128K max output, and is priced at $2 per million input tokens and $10 per million output tokens. OpenAI says it makes about half as many mistakes as GPT-5.6 Sol.
Kimi K3
Open Source
多模态 月之暗面
Kimi K3, released by Moonshot AI on July 16, 2026, is the world's largest open-source model: 2.8T total / 104B activated parameters built on Kimi Delta Attention, Attention Residuals, and Stable LatentMoE (16 of 896 experts per token), with native vision and a 1M-token context window. Optimized for software engineering, knowledge work, deep research, and multimodal tasks, its overall scaling efficiency is ~2.5x that of K2.
Grok 3
Commercial Closed-source
多模态 xAI
Grok 3 is the third-generation large model launched by Musk's xAI, trained on a cluster of 200,000 H100 GPUs. It excels in mathematical reasoning, scientific Q&A, and programming competitions, and has DeepSearch and Think reasoning modes. Grok 3 is designed with the pursuit of truth and geek spirit, scoring over 93% in the AIME math competition, making it one of the most competitive models in AI competitions.
DeepSeek-V3.1
Open Source
文本 DeepSeek
DeepSeek-V3.1 is a major upgrade released by DeepSeek in August 2025, defined as 'the first step towards the Agent era'. It adopts a hybrid reasoning architecture, supporting both thinking and non-thinking modes in one model, with native tool calling and Agent capabilities. It significantly surpasses its predecessor in Agent benchmarks and maintains the MIT open-source license.
Codestral
Open Source
代码 Mistral AI
Codestral is a model specifically designed for code generation by Mistral AI, supporting 80+ programming languages. It excels in code completion, debugging, and refactoring, achieving the highest level among models of the same size on coding benchmarks such as HumanEval and MBPP. Its efficient 22B parameter architecture allows it to run on consumer-grade GPUs, making it an ideal programming assistant for developers.
Qwen2.5-Coder
Open Source
代码 阿里云
Qwen2.5-Coder is an open-source model series specifically designed for coding scenarios by Alibaba Cloud, trained on 5.5 trillion tokens of code data. It comprehensively leads models of the same size on coding benchmarks such as HumanEval, MBPP, and LiveCodeBench, with the 32B version's coding ability even surpassing GPT-4o. Supporting 92 programming languages, it is one of the most powerful code models in the open-source community.
Grok 4.2
Commercial Closed-source
多模态 xAI
Grok 4.2 is the latest large model released by Musk's xAI in February 2026. It has a 2 million token ultra-long context window, supporting ultra-fast inference and ultra-low hallucination. It continues to lead in mathematics, science, and programming competitions, with real-time information retrieval and DeepSearch capabilities. Grok 4.2 inherits the design philosophy of pursuing truth from its predecessor, excelling in complex reasoning and real-time interaction.
Claude Opus 5
Commercial Closed-source
文本 Anthropic
Claude Opus 5 is Anthropic's strongest flagship model released on July 24, 2026. It achieves SOTA on multiple coding evaluations, has a 1 million token context window, and knowledge cutoff of May 2026. Pricing remains at Opus 4.8 levels but performance doubles, making it the current default model for Claude Max. It leads comprehensively in code generation, long-text analysis, and complex reasoning.
Muse Spark 1.3
闭源
多模态 Meta
Muse Spark 1.3, released by Meta on September 2, 2026, is the latest closed-source iteration of the Muse Spark family, focused on strengthening agentic and coding capabilities, powering Meta's personal AI assistant Muse.
GPT-6.1 Sol
Commercial Closed-source
推理 OpenAI
GPT-6.1 Sol is the newest member of OpenAI's GPT-6 family, introduced on September 29, 2026 for hard coding and professional work. OpenAI says it approaches GPT-6 Astra-level capability while its API price is less than half of Astra's standard rate, leaving developers more room to iterate. Specific context window and parameter counts were not published.
Grok 4.7
闭源
推理 xAI
Grok 4.7 is xAI's (now branded SpaceXAI) flagship released on September 21, 2026, targeting coding, long-running agents and knowledge work. It uses a larger base model and a longer reinforcement-learning run, and xAI says it works longer on hard problems, verifies its own output more carefully, and natively understands the Grok Bot harness. The context window stays at 500K tokens with low / medium / high / xhigh reasoning effort, at the same price as Grok 4.6.
Qwen2.5-Math
Open Source
推理 阿里云
Qwen2.5-Math is an open-source model specifically designed for mathematical reasoning by Alibaba Cloud, trained on English and Chinese mathematical data. It achieves industry-leading performance on math benchmarks such as MATH and GSM8K, with the 72B version particularly outstanding on competition-level math problems. It supports two reasoning modes, CoT and TIR (Tool-Integrated Reasoning), making it an ideal tool for math education and research.
MiniMax-Text-01
Commercial Closed-source
文本 MiniMax
MiniMax-Text-01 is an ultra-large-scale MoE model launched by MiniMax, with 456B total parameters and a 1 million token ultra-long context window. It adopts an innovative hybrid mechanism of Lightning Attention and Softmax Attention, providing significant advantages in long document processing. It has reached industry-leading levels on MMLU and Chinese benchmarks, making it an important force in domestic AI models.
Ember-1
Commercial Closed-source
文本 Fireworks AI
Ember-1 is Fireworks Research's first in-house model, released on September 23, 2026 as a Research Preview on serverless, post-trained on Moonshot's open-weight Kimi K3. Fireworks says it roughly matches Kimi K3 quality while emitting 35–50% shorter reasoning (about 40% fewer output tokens), targeting coding, knowledge work and agentic workflows. It is text-in/text-out with a 1,048,576-token context, priced at $3 / $15 per million in/out.
DBRX
Open Source
文本 Databricks
DBRX is an open-source MoE model launched by data platform giant Databricks, adopting a fine-grained expert mixture architecture. It performs excellently in programming, mathematics, and general reasoning, achieving 73.7% accuracy on MMLU. DBRX is a representative work trained by Databricks with an investment of $10 million, showcasing the strength of enterprise-level platforms in the large model field.
Gemini 3.7 Flash
闭源
多模态 Google
Gemini 3.7 Flash, launched by Google on August 13, 2026, is the company's primary coding and agentic model, with major improvements in software engineering, web development, and agentic workflows. It is offered at an introductory price until December 31, 2026.
o4-mini
Commercial Closed-source
推理 OpenAI
o4-mini is an efficient reasoning model released by OpenAI, significantly reducing computational cost while maintaining o3's core reasoning capabilities. It employs an optimized reasoning architecture, performing excellently in mathematics, programming, and scientific tasks, serving as a lightweight alternative to the o3 series. Suitable for scenarios requiring reasoning capabilities but with limited budgets, achieving a new balance in cost-effectiveness.
GPT-4.1-mini
Commercial Closed-source
文本 OpenAI
GPT-4.1-mini is an efficient model launched by OpenAI, significantly reducing cost and latency while maintaining the core capabilities of GPT-4.1. It also has a 1 million token context window and performs close to GPT-4.1 on most tasks, but at only 1/5 of the price. It is the best choice for budget-sensitive scenarios and large-scale deployment.
Claude 4.5 Sonnet
Commercial Closed-source
文本 Anthropic
Claude 4.5 Sonnet is Anthropic's strongest coding model released in 2026, achieving a breakthrough score of approximately 77% on SWE-bench Verified, making it the go-to model for code generation and refactoring. It excels in long document analysis, multi-step reasoning, and complex programming tasks, while maintaining Anthropic's leading position in safety alignment.
DeepSeek V4 Flash
Open Source
文本 DeepSeek
DeepSeek V4 Flash is the latest open-source model released by DeepSeek on July 31, 2026, with a 284B total parameter MoE architecture, activating only 13B parameters for inference. It has a 1 million token context, and API input cost is as low as ¥0.02/million tokens after cache hit. Performance is comparable to GPT-5, Agent capabilities surpass Claude Opus 4.8, and Unsloth released a local running version on the same day. This model was retired on September 10, 2026 with the release of DeepSeek-V4.1-Flash; existing API requests are now served by V4.1-Flash and billed at Flash rates.
GPT-Synopsys
Commercial Closed-source
代码 Upcoming OpenAI
GPT-Synopsys is a chip-design model OpenAI and Synopsys are building under a multi-year agreement, announced on September 30, 2026. It is trained to drive Synopsys EDA tools, read their results and iterate on power, performance and area (PPA). Early customer trials have started, but it has not reached general availability.
未 GA：已公布并开启早期客户试用，尚未正式商用发布
GPT-4.1
Commercial Closed-source
文本 OpenAI
GPT-4.1 is OpenAI's next-generation model built for developers, with significant improvements over GPT-4o in code generation, instruction following, and long-context understanding. It supports a context window of up to 1 million tokens, capable of handling extremely long documents and codebases. It excels in tasks such as frontend code generation, API calls, and structured output.
GPT-6 Astra
闭源
多模态 OpenAI
GPT-6 Astra is OpenAI's flagship model released on September 3, 2026, trained on roughly 100,000 NVIDIA Grace Blackwell NVLink72 GPUs and positioned as the next-generation model for the workplace. It delivers advanced reasoning, computer use, software engineering, cybersecurity, and scientific capabilities, with browser use and multi-step workflows supported across a 1M-token context. API pricing is $10/million input tokens and $50/million output tokens, with a more capable GPT-6 Astra Pro tier.
Gemini Omni 1.1 Flash
闭源
多模态 Google
Gemini Omni 1.1 Flash, released in August 2026 as part of Google's omni-modal world model line, enables agentic video understanding with more control for builders, jointly processing text, image, video, and audio.
Phi-4
Open Source
文本 Microsoft
Phi-4 is a 14B parameter model launched by Microsoft Research, achieving impressive reasoning capabilities with an extremely small parameter size. Trained on high-quality synthetic data, it surpasses many much larger models in mathematical and scientific reasoning. Phi-4 performs outstandingly on GPQA and MATH benchmarks, demonstrating that small models can achieve high-level reasoning through quality data training.
Claude 4 Opus
Commercial Closed-source
文本 Anthropic
Claude 4 Opus is Anthropic's most powerful model, designed for high-difficulty scenarios requiring deep reasoning, complex analysis, and creative writing. It excels in long-document understanding, multi-step reasoning, and cross-domain knowledge integration. Claude 4 Opus achieves top-tier levels on academic benchmarks such as GPQA and MMLU, while maintaining Anthropic's leading advantages in safety and reliability.
Qwen3.8-Flash
Open Source
多模态 阿里云
Qwen3.8-Flash, released by Qwen in August 2026, is a multimodal LLM natively supporting million-level context to process very long documents, code repositories, and complex conversations in one pass. It is compatible with both OpenAI and Anthropic API conventions, seamlessly plugging into Claude Code, Codex, and other developer tools, with strong coding assistance, agentic collaboration, and image-text understanding at competitive inference cost.
Solar Mini 4
Commercial Closed-source
文本 Upstage
Solar Mini 4 is a small, fast agentic model from Korea's Upstage, released on September 23, 2026 and distinct from Solar Pro 4. It is a 35B-total / 3B-active MoE with a 524,288-token context — about a 16x jump over Solar Mini 3 — and up to 128,000 max output, adding reasoning, structured outputs and tool calling. It is proprietary and Korean-optimized multilingual.
DeepSeek-V3
Open Source
文本 DeepSeek
DeepSeek-V3 is DeepSeek's flagship MoE model with a total of 671B parameters, activating only 37B parameters per inference. It uses Multi-head Latent Attention (MLA) and DeepSeekMoE architecture, with a training cost of only $5.57 million, yet achieving performance comparable to GPT-4o and Claude 3.5 Sonnet. It excels particularly in mathematics and coding tasks, and is one of the most influential models in the open-source community.
Gemini 4 Argon
Commercial Closed-source
多模态 Upcoming Google
Gemini 4 Argon is Google DeepMind's new frontier model announced on September 30, 2026 — the first new Gemini generation after the cancelled Gemini 3.5 Pro. It supports a 1M-token context and up to 1M tokens of output, with an introductory price of $2 per million input tokens and $10 per million output tokens (later $4 / $20). It is currently limited to vetted cyber defenders in Google's Fairwind program; no public API has shipped yet.
已公布但未公开发布：仅限 Google Fairwind 网络安全合作方试用，公开 API 与 Google AI Ultra 后续开放
Gemini 3.5 Flash
Commercial Closed-source
多模态 Google
Gemini 3.5 Flash is an efficient multimodal model released by Google at I/O 2026, achieving a new breakthrough in speed and intelligence. It inherits the core capabilities of Gemini 3 Pro and serves with lower latency and cost. It supports full-modal understanding of text, images, video, audio, and code, making it ideal for large-scale deployment of multimodal applications.
DeepSeek-V3.2
Open Source
文本 DeepSeek
DeepSeek-V3.2 is DeepSeek's latest flagship model, with comprehensive upgrades in coding and general capabilities. API pricing is as low as $0.14/M tokens, earning it the title of 'the cheapest top-tier model'. It supports Agent capabilities, long-context understanding, and multilingual dialogue, ranking high on SWE-bench and LMArena.
Jamba 1.5
Open Source
文本 AI21 Labs
Jamba 1.5 is an innovative open-source model launched by AI21 Labs, combining the advantages of the Mamba state space model and Transformer architecture. It has a 256K token ultra-long context window, providing extremely high throughput and low memory usage when processing long documents. Jamba 1.5 achieves a new balance between long-context processing efficiency and reasoning quality, with the Mini version (52B) running on a single GPU.
Clef
Open Source
推理 Cloudflare
Clef is Cloudflare's open decision model released on October 1, 2026, built on Qwen 3.8-27B under the Apache-2.0 license. Instead of writing prose it scores typed choices and returns a probability, accepts the Jev API, and runs on Workers AI — answering a judgment call in milliseconds, which suits routing, ranking and filtering steps.
Llama 3.3
Open Source
文本 Meta
Llama 3.3 is the final version of Meta's Llama 3 series, achieving performance close to Llama 3.1 405B at the 70B parameter scale. It uses a dense architecture (not MoE) and performs excellently in text generation, code writing, and multilingual tasks. Llama 3.3 provides high-quality text generation capabilities at lower deployment costs, and is one of the most popular base models in the open-source community.
SenseNova 5.5
Commercial Closed-source
多模态 商汤科技
SenseNova 5.5 is the latest version of SenseTime's 'Ririxin' large model, with comprehensive upgrades in multimodal understanding and generation. It has unique advantages in image understanding, video analysis, and 3D content generation, deeply integrating SenseTime's computer vision technology accumulation. SenseNova 5.5 is widely applied in vertical scenarios such as autonomous driving, medical imaging, and intelligent security.
Falcon 3
Open Source
文本 TII
Falcon 3 is the third-generation open-source model series launched by TII in Abu Dhabi, trained on 14 trillion tokens. It surpasses other open-source models of the same size, including Llama 3.2 and Qwen 2.5, on benchmarks such as MMLU and HellaSwag. Falcon 3 is released under the Apache 2.0 license, fully open for commercial use, and represents the strength of AI in the Middle East.
CogView-4
Commercial Closed-source
多模态 智谱 AI
CogView-4 is Zhipu AI's latest generation text-to-image model, with major breakthroughs in image quality and text adherence. It supports bilingual prompts in Chinese and English, can generate clear and readable complex text within a single image, and supports arbitrary resolution output. CogView-4 has reached industry-leading levels in image realism, creativity, and prompt understanding, making it one of the strongest text-to-image models in China.
Qwen3.5-Plus
Open Source
文本 阿里云
Qwen3.5-Plus is a new-generation flagship model open-sourced by Alibaba Cloud in February 2026, with a comprehensive architectural overhaul, topping the global list of strongest open-source models. The Qwen series has surpassed 1 billion downloads, overtaking Llama to become the most downloaded open-source model on Hugging Face. It leads comprehensively in coding, mathematics, and general capabilities, supporting thinking mode and Agent tool calling.
Grok 4.6
闭源
多模态 xAI
Grok 4.6, released by xAI on August 12, 2026, is the flagship model of the 1.5T-parameter family, shifting focus to long-horizon agents, coding, knowledge work, and interactive project development. Its overall score matches GPT-5.6 Sol, it ranks first globally on MedAgentBench with 95.9% accuracy, and reaches 67.0% on DeepSWE v1.1 at xhigh thinking effort.
Claude Sonnet 5.5
Commercial Closed-source
文本 Anthropic
Claude Sonnet 5.5 is Anthropic's new-generation Sonnet model released on September 28, 2026, positioned as the best combination of speed and intelligence. It keeps a 1M-token context window and 128K max output, with adaptive thinking on by default (effort defaults to high) so depth can be tuned per task to control cost. Pricing stays at $2 per million input tokens and $10 per million output tokens, unchanged from Sonnet 5, making it the more economical everyday workhorse alongside Opus 5.5.
MiMo-V2.6-Pro
Open Source
多模态 小米
MiMo-V2.6-Pro is Xiaomi's flagship open-weights model released on September 21, 2026, succeeding the proprietary MiMo-V2.5-Pro. It is a 1.02T-total / 42B-active sparse MoE that is natively omni-modal (text, image, video and audio input), with a 1M-token context and 128K max output. Weights are published on Hugging Face under MIT.
DeepSeek-R1
Open Source
推理 DeepSeek
DeepSeek-R1 is a reasoning-enhanced model released by DeepSeek, and the first open-source large reasoning model. It achieves chain-of-thought reasoning capabilities through reinforcement learning training, performing on par with OpenAI o1 in tasks such as mathematics, programming, and scientific reasoning. R1's full open-source caused a stir in the global AI community, proving that open-source models can rival the most advanced commercial closed-source models in reasoning ability.
Llama 4 Scout
Open Source
多模态 Meta
Llama 4 Scout is Meta's lightweight open-source multimodal model, using MoE architecture, with 17B active parameters and an astonishing 10 million token context window. It natively supports text and image understanding and can run on a single H100 GPU. Scout excels in long-document processing and code understanding, and is the choice with the longest context window among open-source models.
Claude Fable 5.1
闭源
多模态 Anthropic
Claude Fable 5.1, released by Anthropic on September 1, 2026, is the company's most advanced model for coding and knowledge work, with long-running agentic coding, knowledge work, and research capabilities. It offers a default 1M-token context window, 128K max output, and always-on adaptive thinking, priced at $10/$50 per million tokens, with prompt cache reads as low as $0.25 per million tokens, making it the top choice for long-horizon agentic tasks.
MiMo-V2.6-Flash
Open Source
多模态 小米
MiMo-V2.6-Flash is the cheaper open-weights sibling of MiMo-V2.6-Pro, released the same day, September 21, 2026. It is a 309B-total / 15B-active MoE (48 layers, 256 experts with top-8 routing, hybrid attention) trained on roughly 48T tokens, natively omni-modal, with a 1M-token context, 128K max output and MIT-licensed weights.
Gemini 3 Pro
Commercial Closed-source
多模态 Google
Gemini 3 Pro is Google's most powerful model released in November 2025, featuring a 2 million token ultra-long context window and native video generation capability. It is the only model that can directly compete with GPT-5.2, leading comprehensively in multimodal understanding, code generation, and scientific reasoning. Gemini 3 Pro marks a new stage in multimodal AI.
Llama 4 Maverick
Open Source
多模态 Meta
Llama 4 Maverick is Meta's high-performance open-source multimodal model, also using MoE architecture, with 400B total parameters and 17B active parameters. It excels in image understanding, coding, and reasoning, with performance comparable to GPT-4o. Maverick has an ultra-long context window of 10 million tokens, providing significant advantages in long-document analysis, codebase understanding, and multimodal dialogue.
Llama 4.1
Open Source
多模态 Meta
Llama 4.1 is Meta's latest generation open-source multimodal model released in 2026, continuing the MoE architecture of the Llama 4 series with comprehensive upgrades. It significantly improves coding, mathematics, and visual understanding, with the context window extended to 10 million tokens. Llama 4.1 strengthens Agent tool calling and code execution capabilities, making it one of the most competitive models in the open-source community.
GPT-4o
Commercial Closed-source
多模态 OpenAI
GPT-4o is OpenAI's flagship multimodal model, natively supporting text, image, and audio input and output. It excels in understanding visual content and voice interaction, with an average response speed 2x faster than GPT-4 Turbo and 50% lower cost. As OpenAI's most versatile model, GPT-4o maintains a leading position in multiple benchmarks such as MMLU and HumanEval.
Ministral 8B
Open Source
文本 Mistral AI
Ministral 8B is Mistral AI's edge model, designed for local devices and offline scenarios. It achieves excellent reasoning capabilities and multilingual support at 8B parameters, running smoothly on laptops and phones. Ministral 8B performs well in knowledge Q&A, translation, and summarization tasks, making it one of the best choices for edge AI deployment.
Llama 3.2
Open Source
多模态 Meta
Llama 3.2 is a multimodal model series launched by Meta, introducing visual understanding capabilities for the first time in the Llama series. It offers lightweight text-only models of 1B and 3B, and multimodal models of 11B and 90B, covering deployment scenarios from phones to data centers. The 90B version excels in image understanding, document analysis, and multimodal dialogue.
Gemma 3
Open Source
多模态 Google
Gemma 3 is an open-source lightweight model series released by Google, built on Gemini 2.0 technology. It offers four sizes from 1B to 27B, supports text and vision multimodal input, and has a 128K token context window. Gemma 3 leads in performance among models of the same parameter size, can run on a single GPU or even a phone, and is a high-quality choice for fine-tuning and deployment in the open-source community.
GPT-5
Commercial Closed-source
多模态 OpenAI
GPT-5 is OpenAI's flagship large model launched in August 2025, natively supporting Agent capabilities and tool calling. It features deep reasoning, multimodal understanding (text/image/audio/video), and code generation, setting new records on benchmarks such as SWE-bench and GPQA. GPT-5's Agent Plugins specification has become the standard for AI agent plugin packaging.
Clef-flash
Open Source
推理 Cloudflare
Clef-flash is Cloudflare's smaller open decision model, released on October 1, 2026 alongside Clef, built on Qwen 3.5-9B under Apache-2.0. It uses the same Jev API and likewise returns probability scores for given options, with lower latency and cost for high-concurrency, lightweight decisions.
Step-2
Commercial Closed-source
多模态 阶跃星辰
Step-2 is a trillion-parameter MoE multimodal large model launched by StepFun, excelling in mathematics, logical reasoning, and image understanding. It adopts an innovative multimodal fusion architecture and performs well in image generation, video understanding, and code writing. Step-2 has reached industry-leading levels on multiple international benchmarks, making it an important force in the domestic multimodal AI field.
ERNIE Bot 4.0 Turbo
Commercial Closed-source
多模态 百度
ERNIE Bot 4.0 Turbo is Baidu's flagship large model, optimized based on the ERINE 4.0 architecture. It has four core capabilities: understanding, generation, logic, and memory, performing particularly well in Chinese scenarios. It supports various tasks such as text generation, image understanding, and code writing, and provides enterprise-level API services through Baidu AI Cloud Qianfan platform, making it one of the most widely used large models in China.
Qwen2.5-VL
Open Source
多模态 阿里云
Qwen2.5-VL is a vision-language model from Alibaba's Tongyi Qianwen team, supporting deep understanding of images, videos, and documents. It can process images of any size with dynamic resolution, support video analysis of up to several hours, and accurately locate object coordinates in images. The 72B version reaches industry-leading levels in multimodal benchmarks, and is one of the benchmarks for open-source vision-language models.
Hunyuan-Large
Open Source
文本 腾讯
Hunyuan-Large is an open-source MoE model launched by Tencent Hunyuan team, with 389B total parameters and 52B active parameters. It is one of the largest open-source MoE models currently, performing excellently on benchmarks such as MMLU and CMMLU. Hunyuan-Large performs particularly well in Chinese scenarios, with outstanding performance in long text processing, knowledge Q&A, and code generation.
Command R7B
Open Source
文本 Cohere
Command R7B is a lightweight open-source RAG model launched by Cohere, achieving excellent retrieval-augmented generation capabilities at 7B parameters. It supports multilingual processing in 23 languages, has built-in citation functionality, and can run efficiently on consumer-grade GPUs. Command R7B brings enterprise-level RAG capabilities to small models and edge devices, making it a popular choice for RAG application development.
Qwen3.8-Max
Open Source
多模态 阿里云
Qwen3.8-Max, the open flagship released by Alibaba's Qwen team in August 2026, is a sparse MoE with 2.4T total parameters (~95B activated per step), hybrid attention, 1M-token context, and 131K output. It excels at complex engineering-level coding and long-horizon autonomous development, scoring GPQA Diamond 92.6, PaperBench 93.0, OSWorld 86.1, and BabyVision 82.0 (CodeArena #4 globally). Pricing: ¥12/M input, ¥36/M output (¥1.5/M cache hit).
GPT-5.2 Instant
Commercial Closed-source
文本 OpenAI
GPT-5.2 Instant is the efficient version of the GPT-5.2 series, achieving the best balance between speed and quality. It inherits the core capabilities of GPT-5.2 Pro and serves with lower latency and cost, making it ideal for production environments and large-scale deployment. It supports Agent plugins and tool calling.
Pixtral Large
Open Source
多模态 Mistral AI
Pixtral Large is Mistral AI's flagship vision-language model, based on the Mistral Large 2 architecture with a new advanced vision encoder. It excels in document understanding, chart analysis, and multimodal reasoning, supporting a 128K token context and multilingual processing. Pixtral Large is a leader among open-source vision-language models and can handle complex mixed text-image tasks.
Gemini 2.5 Flash
Commercial Closed-source
多模态 Google
Gemini 2.5 Flash is Google's efficient thinking model, achieving an optimized balance between speed and intelligence. It also supports a 1 million token context and native multimodality, and has thinking and reasoning capabilities. Compared to the Pro version, Flash provides similar intelligence with lower latency and cost, making it an ideal choice for most production scenarios.
GLM-5.3-Flash
Open Source
多模态 智谱 AI
GLM-5.3-Flash, released by Zhipu in August 2026, is the first natively multimodal model of the GLM-5 series: 320B total / 18B activated parameters with a hybrid sparse+linear attention architecture, delivering intelligence beyond GLM-5.2 at lower cost. It supports 1M context and 128K output, native image/video/file input, and can coordinate across code, browser, and GUI environments.
Mistral Large 2
Commercial Closed-source
文本 Mistral AI
Mistral Large 2 is Mistral AI's flagship model with 123B parameters, excelling in code generation, mathematical reasoning, and multilingual processing. It supports dozens of natural languages and over 80 programming languages, achieving 84.0% accuracy on MMLU. Mistral Large 2 achieves performance close to GPT-4o with a relatively small parameter size, and is an outstanding representative of European AI.
DeepSeek-Janus-Pro
Open Source
多模态 DeepSeek
DeepSeek-Janus-Pro is a unified multimodal understanding and generation model launched by DeepSeek. It adopts an innovative decoupled visual encoding architecture, supporting both image understanding (description, Q&A, OCR) and image generation (text-to-image). The 7B parameter design enables efficient operation, opening a new technical path in unified multimodal models.
Claude Opus 5.5
Commercial Closed-source
推理 Anthropic
Claude Opus 5.5 is Anthropic's new frontier model released on September 22, 2026, built for long-running agentic coding and knowledge work. It reaches roughly Claude Fable 5.1-level quality on most tasks while cutting running cost about 40% below Opus 5: input and output prices are 20% lower and cache reads 60% cheaper. It offers a 1M-token context, 128K max output (up to 300K with the Batch API), and a June 2026 knowledge cutoff.
Spark 4.0
Commercial Closed-source
多模态 科大讯飞
Spark Cognitive Large Model 4.0 is the latest version launched by iFlytek, with unique advantages in voice interaction, multimodal understanding, and educational scenarios. It deeply integrates iFlytek's technical accumulation in speech recognition and synthesis, supporting high-quality voice interaction in 37 languages. Spark 4.0 has rich application practices in vertical industries such as education, healthcare, and law.
DeepSeek-V4.1-Flash
Open Source
多模态 DeepSeek
DeepSeek-V4.1-Flash is the smallest member of DeepSeek's new architecture family, released on September 10, 2026. It is a 552B-parameter MoE model built on the new Causal Encoder-Decoder (CED) asymmetric architecture, activating only 8B parameters for input and 16B for output, designed for the 'input-heavy, output-light' workloads of the Agent era. It supports a 1M-token context with up to 384K output, native multimodal vision understanding, and a KV cache that needs just 1/4 of the HBM of the previous generation. Its benchmarks comprehensively surpass the flagship V4 Pro, and API pricing is as low as ¥0.02/million tokens on cache hits, making it the most cost-effective open-source model of the Agent era.
GLM-4-Plus
Commercial Closed-source
多模态 智谱 AI
GLM-4-Plus is a flagship large model launched by Zhipu AI, adopting a MoE architecture, with comprehensive upgrades in language understanding, code generation, and image understanding. It supports All-Tools mode, which can automatically invoke tools such as search, code interpreter, and drawing, and has strong Agent capabilities. GLM-4-Plus performs exceptionally well on Chinese tasks and is an important choice for enterprise-level AI applications in China.
Phi-4-multimodal
Open Source
多模态 Microsoft
Phi-4-multimodal is an ultra-lightweight multimodal model launched by Microsoft, with only 5.6B parameters yet supporting full modality processing of speech, vision, and text. It can run in real-time on mobile devices, supporting speech recognition, image understanding, and text generation. Phi-4-multimodal sets a new benchmark in small-model multimodal field, making it the preferred multimodal solution for edge AI devices.
GPT-6 Luna
Commercial Closed-source
文本 OpenAI
GPT-6 Luna is the lower-tier, high-throughput member of the GPT-6 family, released on September 22, 2026 alongside GPT-6 Sol. It keeps the same ~1.05M-token context and 128K max output but pushes pricing down to $0.10 per million input tokens and $0.50 per million output tokens — roughly half or less of the previous tier — for large volumes of everyday tasks.
Gemini 4 Pro
Commercial Closed-source
多模态 Upcoming Google
Gemini 4 Pro is the flagship tier of Google DeepMind's next Gemini generation and has not been officially released. Google confirmed in its July 2026 earnings call that it had started its "most ambitious pretraining run yet", and DeepMind's lead said on September 24 that Gemini 4 is in post-training and Google wants to ship "as soon as possible". A leaked internal checkpoint (codename argon) is reported to allow 256K output tokens and possibly target a 1.5M+ context.
未发布：Google 已确认 Gemini 4 在训练 / 后训练中，非官方时间表指向 2026 年 10 月
DeepSeek-Coder-V2
Open Source
代码 DeepSeek
DeepSeek-Coder-V2 is an open-source model specifically designed for coding scenarios by DeepSeek, based on MoE architecture, trained on 338 programming languages. It surpasses GPT-4 Turbo on coding benchmarks such as HumanEval and MBPP, and performs excellently in code completion, debugging, and refactoring. Supports 128K context, capable of handling large codebases, making it an ideal programming assistant for developers.
DeepSeek-V2.5
Open Source
文本 DeepSeek
DeepSeek-V2.5 is a major upgrade of DeepSeek-V2, integrating the code capabilities of DeepSeek-Coder-V2 and the general capabilities of DeepSeek-V2. It achieves industry-leading performance in mathematics, programming, and general tasks, and supports practical features such as web search and file upload. V2.5 provides performance close to GPT-4o at extremely low deployment cost.
Space Bunny
闭源
多模态 未公开
Space Bunny is an anonymous Flash-class large model that Tencent's WorkBuddy announced as an exclusive domestic integration on October 2, 2026 (with a limited-time discount from October 2 to 7). It is described as offering very fast reasoning and strong coding, natively supporting text, image and video input, adjustable thinking depth and a 1M-token context. IT Home reported it ranked first in token usage on both OpenRouter and OpenCode for three consecutive days.
GPT-5.2 Pro
Commercial Closed-source
推理 OpenAI
GPT-5.2 Pro is OpenAI's most powerful reasoning model, released in December 2025. It excels in professional knowledge work (spreadsheets, presentations, complex reports) and can handle multi-step complex reasoning in deep thinking mode. Together with Instant and Thinking, it forms the three variants of the GPT-5.2 series.
Claude 3.5 Sonnet
Commercial Closed-source
文本 Anthropic
Claude 3.5 Sonnet is Anthropic's mid-range flagship model, achieving the best balance between intelligence and speed. It excels in programming, creative writing, and complex analysis, supporting a 200K token context. It achieved breakthrough results on the SWE-bench coding benchmark while maintaining Anthropic's leading standards in safety and reliability.
Kimi K2
Commercial Closed-source
文本 月之暗面
Kimi K2 is an ultra-large-scale MoE model released by Moonshot AI, with a total parameter count exceeding 1 trillion, making it one of the largest language models currently available. It excels in long-text understanding, deep reasoning, and knowledge Q&A, inheriting the Kimi series' consistent advantage in long-context processing. Kimi K2 has unique advantages in professional domain knowledge Q&A and complex document analysis scenarios.
GPT-5.2 Thinking
Commercial Closed-source
推理 OpenAI
GPT-5.2 Thinking is the thinking-enhanced version of the GPT-5.2 series, designed for mathematics, programming, and scientific scenarios that require deep reasoning. While maintaining the core capabilities of GPT-5.2, it achieves deeper analysis through an enhanced chain-of-thought reasoning mechanism, making it the most cost-effective reasoning model.
GPT-5.6 Sol
Commercial Closed-source
推理 OpenAI
GPT-5.6 Sol is OpenAI's current strongest flagship reasoning model, released in July 2026. The GPT-5.6 series has three tiers: Luna (entry-level free), Terra (mid-range), and Sol (flagship). Sol unifies Instant and Thinking modes through an innovative slider mechanism, comprehensively refreshing SOTA in coding, reasoning, and multimodal tasks.
Claude Opus 4.8
Commercial Closed-source
文本 Anthropic
Claude Opus 4.8 is a flagship model released by Anthropic in May 2026, just 41 days after the previous version Opus 4.7—setting the shortest iteration record for the Claude flagship series. It comprehensively upgrades coding, reasoning, and safety alignment, laying the technical foundation for the subsequent Claude Science research platform and Claude Mythos 5.
Gemini 2.5 Pro (Thinking)
Commercial Closed-source
推理 Google
Gemini 2.5 Pro's enhanced thinking version performs deeper chain-of-thought reasoning before answering. It particularly strengthens mathematical, scientific, and logical reasoning abilities, significantly outperforming the regular version in complex problems requiring multi-step derivation. It supports a 1 million token context and native multimodal input, making it Google's most powerful reasoning model.
Doubao Doubao-1.5-pro
Commercial Closed-source
文本 字节跳动
Doubao Doubao-1.5-pro is ByteDance's flagship large language model, using a large-scale MoE architecture, excelling in Chinese language understanding, knowledge Q&A, and creative writing. It has an ultra-long context window of 256K tokens, providing significant advantages in long-document processing and multi-turn dialogue scenarios. It is provided through the Doubao platform and Volcano Engine, deeply integrated with ByteDance's product ecosystem.
Claude Mythos 5.1
闭源
多模态 Anthropic
Claude Mythos 5.1, released on September 1, 2026, shares the same foundation as Claude Fable 5.1 with the strongest coding and knowledge-work capabilities, but uses a different safeguard profile (retaining cybersecurity and bioscience restrictions). It is available only through the Project Glasswing trusted-partner program for highly sensitive environments.
Qwen3.8-27B
Open Source
多模态 阿里云
Qwen3.8-27B, open-sourced by Alibaba on August 14, 2026, is a natively multimodal dense model with 262K native context (extendable to 1M via YaRN) under the Apache 2.0 license. It scores SWE-bench Pro 61.7, LiveCodeBench v6 90.3, GPQA Diamond 89.2, and OSWorld-Verified 84.3, striking the golden balance between performance and cost for low-barrier deployment.
GPT-4o-mini
Commercial Closed-source
文本 OpenAI
GPT-4o-mini is OpenAI's most cost-effective small model, designed to provide GPT-4o-level core capabilities at a very low price. It supports text and vision input, scores 82% on MMLU, and surpasses GPT-3.5 Turbo in mathematical reasoning and coding tasks. Suitable for simple task processing in large-scale production environments.
Yi-Large
Commercial Closed-source
文本 零一万物
Yi-Large is the flagship model launched by 01.AI, benchmarked against GPT-4.0 in comprehensive capabilities. It has an ultra-long context window of 200K tokens and excels in Chinese understanding, knowledge Q&A, and logical reasoning. Yi-Large achieves 88.5% accuracy on MMLU, making it one of the earliest domestic models to reach GPT-4 level performance.
Qwen2.5-Max
Commercial Closed-source
文本 阿里云
Qwen2.5-Max is Alibaba Cloud's flagship commercial closed-source model, using a large-scale MoE architecture, benchmarking against GPT-4o and Claude 3.5 Sonnet in overall capabilities. It excels in Chinese understanding, knowledge Q&A, and complex reasoning, with special optimization for mixed Chinese-English scenarios. It is provided through Alibaba Cloud Bailian platform, targeting enterprise-level customers.
Hunyuan Turbo
Commercial Closed-source
多模态 腾讯
Hunyuan Turbo is Tencent Hunyuan's latest flagship commercial closed-source model, based on a large-scale MoE architecture, with comprehensive upgrades in inference speed and intelligence. It supports native multimodal (text, image, video understanding), has a 256K token ultra-long context, and excels in programming, mathematics, and creative writing. Deeply integrated with WeChat and Tencent ecosystem, it is one of the largest AI models by user base in China.
Command R+
Commercial Closed-source
文本 Cohere
Command R+ is Cohere's flagship enterprise-grade model, optimized for RAG (Retrieval-Augmented Generation) and tool use scenarios. It excels in multilingual processing, long document analysis, and enterprise workflows. Command R+ has a built-in powerful citation feature that automatically provides source annotations for generated content, making it an ideal choice for enterprise AI applications.
Perceptron Mk1.5
Commercial Closed-source
多模态 Perceptron
Perceptron Mk1.5 is a hosted proprietary embodied-reasoning and perception model released on September 25, 2026 for physical agents, succeeding Mk1 and retiring the open-weight isaac-0.1/0.2 checkpoints. It accepts text, image, video and audio input and returns natural-language text with optional machine-parseable annotations (points, boxes, polygons, tracks, clips), with reasoning-effort control, tool calling and JSON-Schema structured outputs. Context is 36,864 tokens with 8,192 max output.
GPT-4.1-nano
Commercial Closed-source
文本 OpenAI
GPT-4.1-nano is OpenAI's fastest and most economical model, designed for simple tasks and ultra-high throughput scenarios. It has a 1 million token context window and performs well in classification, extraction, and simple Q&A tasks. With extremely low latency and a price only 1/20 of GPT-4.1, it is ideal for automated pipelines and large-scale data processing.
没有匹配的模型，试试更换筛选条件。
