GPT, Claude, Gemini et Grok sont de bons choix si vous avez besoin de raisonnement general, de code, de redaction, de vision et d outils dans un seul assistant.
Repertoire des modeles IA
Comparez les modeles IA qui transforment le travail, la recherche, le code, l image et la video.
Utilisez ce repertoire pour comprendre qui cree les principaux modeles IA, d ou ils viennent, quelle est leur taille lorsque l information est publique, a quoi ils servent et lesquels meritent une analyse pour les workflows d entreprise.
Lecture rapide
Les familles Kimi, Qwen, Llama, DeepSeek et GLM montrent pourquoi les modeles mixture-of-experts comptent : tres grande taille totale avec moins de parametres actifs par token.
Nano Banana, Veo, Sora, Midjourney, Flux, Runway, Kling et les modeles proches se jugent davantage sur le controle, la coherence et les droits que sur le nombre de parametres.
Les modeles ouverts plus petits peuvent etre utiles pour les workflows prives, l aide au code et l experimentation lorsque le meilleur modele est celui que vous pouvez executer et gouverner.
Plus grandes familles de modeles publiques ou rapportees
MoE signifie mixture-of-experts. Les parametres totaux decrivent le modele complet; les parametres actifs decrivent la partie utilisee pour un token.
Announced around 2T total parameters, 288B active. Not generally released.
1T total, 32B active. Open-weight agentic and coding-focused MoE family.
Reported above 1T parameters. Proprietary API model from Alibaba/Qwen.
Widely reported 671B total, 37B active for major open MoE releases.
400B total, 17B active. Open multimodal MoE model family.
Tableau comparatif des modeles IA
Tous les modeles sont affiches
Taille affichee lorsqu elle est publique
Mis a jour le 6 juillet 2026
Aucun modele ne correspond aux filtres actuels.
| Model | Fabricant | Pays / region | Modalite | Taille / contexte | Acces | Ideal pour | Notes pratiques |
|---|---|---|---|---|---|---|---|
| GPT-5.6 PreviewOpenAI | OpenAI | United States | Multimodal Code | Parameter count not published. Preview availability. GPT-5.5 family lists 1M context and 128K max output. | API preview for selected partners | Advanced reasoning, coding and professional work. | Use when frontier capability matters more than predictable cost or broad availability. |
| GPT-5.5OpenAI flagship model | OpenAI | United States | Multimodal Code | Parameter count not published. 1M context, 128K max output in OpenAI API docs. | API and ChatGPT ecosystem | Complex reasoning, coding, agents, writing, analysis and computer-use workflows. | Strong default for teams that need one high-capability assistant across many tasks. |
| GPT-5.4OpenAI | OpenAI | United States | Multimodal | Parameter count not published. 1M context, 128K max output. | API | Professional writing, coding and lower-cost reasoning compared with GPT-5.5. | Useful when you need strong output but want a more controlled budget. |
| GPT-5.4 miniOpenAI | OpenAI | United States | Multimodal | Parameter count not published. 400K context, 128K max output. | API | High-volume coding, subagents and practical automation. | Good fit for workflows where latency and cost matter. |
| GPT Image 2OpenAI image generation | OpenAI | United States | Image | Parameter count not published. | API | Image generation and editing for product, marketing and creative workflows. | Evaluate on brand consistency, text rendering, edit control and usage rights. |
| GPT-Realtime-2OpenAI realtime voice | OpenAI | United States | Audio | Parameter count not published. | API | Realtime voice agents, spoken interfaces and translation workflows. | Measure interruption handling, latency, cost per conversation and safety controls. |
| Claude Fable 5Anthropic | Anthropic | United States | Multimodal | Parameter count not published. 1M context, 128K max output. | Claude API, AWS, Google Cloud, Microsoft Foundry | Long-running agents, enterprise reasoning and high-stakes structured work. | Consider when long context and careful outputs are central to the workflow. |
| Claude Opus 4.8Anthropic | Anthropic | United States | Multimodal Code | Parameter count not published. 1M context, 128K max output. | Claude API and cloud platforms | Complex agentic coding, long documents, analysis and enterprise work. | Useful for careful writing, review and multi-step work with large context. |
| Claude Sonnet 5Anthropic | Anthropic | United States | Multimodal Code | Parameter count not published. 1M context, 128K max output. | Claude API, Claude app and cloud platforms | Balanced speed, intelligence, coding, writing and analysis. | Often a practical everyday model when quality and speed both matter. |
| Claude Haiku 4.5Anthropic | Anthropic | United States | Multimodal | Parameter count not published. 200K context, 64K max output. | Claude API and cloud platforms | Fast, lower-cost responses for near-frontier quality tasks. | Good for classification, extraction, support drafts and lightweight automations. |
| Gemini 2.5 ProGoogle DeepMind | Google DeepMind | United States / United Kingdom | Multimodal Code | Parameter count not published. | Gemini API, Google AI Studio, Vertex AI, Gemini app | Complex reasoning, coding and multimodal workflows. | Strong fit when your team already works in Google Workspace and Vertex AI. |
| Gemini 2.5 FlashGoogle DeepMind | Google DeepMind | United States / United Kingdom | Multimodal | Parameter count not published. | Gemini API and Gemini app | Low-latency, high-volume multimodal tasks that still need reasoning. | Use for assistants, extraction, summaries and scalable workflow support. |
| Gemini 2.5 Flash-LiteGoogle DeepMind | Google DeepMind | United States / United Kingdom | Multimodal | Parameter count not published. | API | Fast and cost-sensitive multimodal production tasks. | Best where throughput matters more than maximum reasoning depth. |
| Nano Banana 2Gemini 3.1 Flash Image | Google DeepMind | United States / United Kingdom | Image | Parameter count not published. | Gemini app, API and Vertex AI | Fast image generation and editing with strong instruction following. | Useful for quick creative iteration, product images and social visuals. |
| Nano Banana ProGemini 3 Pro Image | Google DeepMind | United States / United Kingdom | Image | Parameter count not published. Supports studio-quality visual workflows and 4K-oriented output according to Google docs. | Gemini app, API and Vertex AI | Higher-end design, text rendering and complex image edits. | Compare against Midjourney, Flux and Ideogram for brand image workflows. |
| Nano Banana LiteGoogle DeepMind | Google DeepMind | United States / United Kingdom | Image | Parameter count not published. | API | Low-latency image generation and editing. | Designed for cost-sensitive production workflows. |
| Veo 3.1 PreviewGoogle DeepMind | Google DeepMind | United States / United Kingdom | Video | Parameter count not published. | API preview and Google creative surfaces | Cinematic video generation with creative controls and synchronized audio. | Evaluate shot consistency, brand safety, editability and rights before production use. |
| Gemini 3.1 Flash Live PreviewGoogle DeepMind | Google DeepMind | United States / United Kingdom | Audio | Parameter count not published. | API preview | Realtime voice-first AI applications. | Useful for customer support, tutoring and voice interfaces when latency is critical. |
| Grok 4.3xAI | xAI | United States | Text | Parameter count not published. 1M context in xAI docs. | xAI API and Grok products | General reasoning, current-style conversational work and broad assistant tasks. | Use search tools when current events or fresh web context matter. |
| Grok Build 0.1xAI coding model | xAI | United States | Code | Parameter count not published. | xAI API | Agentic coding workflows. | Choose for code-specific tasks instead of the general Grok model. |
| Grok ImaginexAI media API | xAI | United States | Image Video | Parameter count not published. | xAI API | Image and video generation inside the Grok ecosystem. | Compare quality and content controls with dedicated media platforms. |
| Mistral Medium 3.5Mistral AI | Mistral AI | France | Multimodal Code | Parameter count not published. | API and open model availability where offered | Agentic and coding use cases with European AI governance needs. | Strong option when EU supplier strategy matters. |
| Mistral Small 4Mistral AI | Mistral AI | France | Multimodal | Parameter count not published. | API and open-weight family | Hybrid instruct, reasoning and coding in efficient workflows. | Good candidate for teams balancing performance, portability and cost. |
| Mistral Large 3Mistral AI | Mistral AI | France | Multimodal | Parameter count not published in overview. | API and open-weight availability where offered | General-purpose multimodal enterprise work. | Consider when an EU-headquartered frontier supplier is preferred. |
| Ministral 3 14BMistral AI | Mistral AI | France | Multimodal | 14B parameters. | API, open-weight and local deployment options | Efficient text and vision use cases. | Practical for smaller deployments and controlled environments. |
| Devstral 2Mistral AI | Mistral AI | France | Code | Parameter count not published in overview. | API and open model availability where offered | Software engineering agents and repository tasks. | Compare against Cursor-specialized and Grok Build-style coding models. |
| CodestralMistral AI | Mistral AI | France | Code | Parameter count not published in overview. | API | Code completion and developer assistance. | Use where autocomplete, code transformation and repository context matter. |
| Voxtral SmallMistral AI | Mistral AI | France | Audio | Parameter count not published in overview. | API and open-weight family | Audio input and speech-oriented instruction use cases. | Useful for transcription, voice workflows and multilingual audio tasks. |
| OCR 4Mistral AI | Mistral AI | France | OCR | Parameter count not published. | API | Document AI, structured OCR and paragraph-level extraction. | Important for businesses with PDF, invoice, archive and compliance workflows. |
| DeepSeek V4 ProDeepSeek | DeepSeek | China | Text Code | 1M context and 384K max output in DeepSeek API docs. Parameter count not published on pricing page. | API | Reasoning, coding, JSON output, tool calls and long-context tasks. | Check data handling, jurisdiction and vendor risk before enterprise use. |
| DeepSeek V4 FlashDeepSeek | DeepSeek | China | Text Code | 1M context and 384K max output in DeepSeek API docs. | API | Low-cost high-volume reasoning and coding tasks. | Likely attractive where token cost is a major constraint. |
| DeepSeek R1Reasoning model lineage | DeepSeek | China | Text Code | Widely reported MoE scale: 671B total, 37B active. | Open-weight and hosted derivatives | Reasoning, math, coding and transparent local experimentation. | Distilled variants are often easier to run than the full model. |
| DeepSeek V3Open MoE model lineage | DeepSeek | China | Text Code | Widely reported MoE scale: 671B total, 37B active. | Open-weight and hosted derivatives | General language, coding and cost-efficient inference. | Important reference point for modern open MoE architecture. |
| Qwen3.5Alibaba Qwen | Alibaba / Qwen | China | Multimodal Code | Family listed from 0.8B to 122B in local model catalogs. | API, open-weight variants and local deployment | Multilingual, multimodal, coding and agentic workflows. | Strong family to watch for local and enterprise model stacks. |
| Qwen3 235B-A22BAlibaba Qwen MoE | Alibaba / Qwen | China | Multimodal Code | 235B total, 22B active in public Qwen3 model family reporting. | Open-weight family | Reasoning, multilingual tasks, coding and high-capability local deployments. | Large enough for serious capability, but much lighter active compute than dense 235B. |
| Qwen3-VL 235B-A22BAlibaba Qwen vision-language | Alibaba / Qwen | China | Multimodal | 235B total, 22B active for MoE variant; 256K-token native context reported. | Open-weight family | Image, video, document and multimodal reasoning. | Useful benchmark model for visual reasoning and document-heavy workflows. |
| Qwen3-MaxAlibaba Qwen proprietary | Alibaba / Qwen | China | Multimodal | Reported above 1T parameters. | API | Large-scale reasoning, media-capable generation and enterprise workflows. | Evaluate against GPT, Claude and Gemini depending on region and compliance needs. |
| Llama 4 MaverickMeta | Meta | United States | Multimodal | 400B total, 17B active; 1M context reported for release family. | Open model family | Open multimodal work, enterprise experimentation and custom deployments. | One of the most important open model families for businesses wanting control. |
| Llama 4 ScoutMeta | Meta | United States | Multimodal | 109B total, 17B active; 10M context reported. | Open model family | Long-context multimodal experimentation and local/cloud deployments. | Interesting for retrieval and large-document workflows where context length matters. |
| Llama 4 BehemothMeta announced model | Meta | United States | Multimodal | Announced around 2T total, 288B active. Not generally released. | Announced / not generally available | Reference point for very large MoE model scale. | Do not plan production workflows around it until availability is clear. |
| Kimi K2Moonshot AI | Moonshot AI | China | Text Code | 1T total, 32B active; trained on 15.5T tokens in published research. | Open-weight and hosted Kimi ecosystem | Agentic coding, long-horizon tasks and software engineering. | One of the largest practical open-weight MoE families. |
| Kimi K2.7 CodeMoonshot AI | Moonshot AI | China | Code | Built on the Kimi K2 lineage; parameter details vary by release. | Hosted and local catalog availability | Long-horizon agentic coding with lower thinking-token usage. | Useful candidate for comparing coding agents against Cursor-focused models. |
| GLM-4.5Z.ai | Z.ai | China | Text Code | 355B total, 32B active; trained on 23T tokens in published research. | Open-weight | Agentic, reasoning and coding workflows. | Strong public model for teams evaluating open MoE reasoning systems. |
| GLM-4.5-AirZ.ai | Z.ai | China | Text Code | 106B total parameters. | Open-weight | More efficient agentic reasoning and coding. | Consider when GLM-4.5 capability is attractive but full-size deployment is too heavy. |
| GLM-5.2Z.ai | Z.ai | China | Code | Parameter count not published in local catalog excerpt. | Cloud and local catalog availability | Long-horizon agentic engineering. | Track closely if your priority is software engineering automation. |
| GLM-OCRZ.ai | Z.ai | China | OCR | Parameter count not published in local catalog excerpt. | Local and cloud catalog availability | Complex document understanding and OCR. | Useful for invoices, forms, PDFs and scanned business records. |
| Gemma 4Google DeepMind open family | Google DeepMind | United States / United Kingdom | Multimodal | Local catalog lists E2B, E4B, 12B, 26B and 31B variants. | Open-weight and local deployment | Efficient reasoning, coding, multimodal and audio-capable workflows. | Good family for privacy-aware deployments and experimentation. |
| Laguna XS 2.1Local agentic coding model | Laguna / local model community | United States / open ecosystem | Code | 33B total MoE, 3B active per token. | Local model catalogs | Agentic coding and long-horizon local work. | Interesting when you want capable coding support on local hardware. |
| North Mini Code 1.0Cohere | Cohere | Canada | Code | 30B total MoE, 3B active parameters. | Local model catalogs and Cohere ecosystem | Agentic software engineering. | Useful for comparing smaller active-parameter coding models. |
| Command ACohere | Cohere | Canada | Text | Commonly reported around 111B parameters. | API | Enterprise search, RAG, multilingual assistants and business workflows. | Often evaluated where retrieval and enterprise controls are central. |
| DBRXDatabricks | Databricks | United States | Text | 132B total MoE, 36B active widely reported for the release. | Open-weight | Enterprise open model experimentation and data-platform workflows. | Still a useful reference for open MoE model architecture. |
| Falcon 180BTechnology Innovation Institute | TII | United Arab Emirates | Text | 180B dense parameters. | Open-weight | Large open language model evaluation and self-hosting. | Compute-heavy compared with modern smaller active-parameter MoE models. |
| Jais 70BCore42 / MBZUAI / Cerebras | Core42 and partners | United Arab Emirates | Text | 70B family. | Open-weight and hosted variants | Arabic and English language workflows. | Important if Arabic-language performance matters. |
| Jamba 1.5 LargeAI21 Labs | AI21 Labs | Israel | Text | Hybrid architecture; parameter details vary by release. | API and open model availability where offered | Long-context enterprise language tasks. | Worth evaluating for structured business writing and long-document workflows. |
| Midjourney v7Midjourney | Midjourney | United States | Image | Parameter count not published. | Web app and platform access | High-quality artistic and commercial image generation. | Strong creative output; check workflow fit, editing control and licensing terms. |
| FLUX.1 familyBlack Forest Labs | Black Forest Labs | Germany | Image | Parameter count varies by model; several public and hosted variants. | API, hosted apps and open-weight variants | Image generation, editing and controllable creative pipelines. | Important European image-model family for brand and product workflows. |
| Stable Diffusion 3.5Stability AI | Stability AI | United States / United Kingdom | Image | Model sizes vary by variant. | API, hosted and open-weight variants | Image generation, customization and creative tooling. | Strong ecosystem with many third-party tools and fine-tuning workflows. |
| Ideogram 3.0Ideogram | Ideogram | Canada / United States | Image | Parameter count not published. | Web app and API availability | Image generation with strong text-in-image use cases. | Useful for ads, posters, labels and visuals where readable text matters. |
| Leonardo PhoenixLeonardo.Ai | Leonardo.Ai | Australia | Image | Parameter count not published. | Web app and creative platform | Marketing images, game assets, concept art and creative workflows. | Good fit when a full creative production interface matters. |
| SoraOpenAI video model | OpenAI | United States | Video | Parameter count not published. | OpenAI creative products and availability by plan/region | Text-to-video and creative video generation. | Evaluate story control, consistency, safety limits and commercial rights. |
| Runway Gen-4Runway | Runway | United States | Video | Parameter count not published. | Creative platform and API availability | Video generation, editing and creative production workflows. | Useful for marketing teams and studios that need controllable video workflows. |
| KlingKuaishou | Kuaishou | China | Video | Parameter count not published. | Web app and API availability | Text-to-video, image-to-video and cinematic motion generation. | Compare motion quality and regional data terms before business use. |
| Dream MachineLuma AI | Luma AI | United States | Video | Parameter count not published. | Creative platform and API availability | Image-to-video and cinematic video generation. | Good for fast concepting and social video experiments. |
| Pika 2.xPika | Pika | United States | Video | Parameter count not published. | Creative platform | Short-form video generation, edits and social creative. | Best evaluated by ease of use and repeatable brand workflows. |
| Hailuo / MiniMax VideoMiniMax | MiniMax | China | Video | Parameter count not published. | Creative platform and API availability | Video generation and consumer creative workflows. | Assess output rights, data handling and regional availability. |
| HunyuanVideoTencent | Tencent | China | Video | Model sizes vary by release. | Open-weight variants and hosted ecosystem | Video generation research and customizable media pipelines. | Relevant when open video-model workflows matter more than managed apps. |
Derniere revision : 6 juillet 2026. La disponibilite, les noms, les prix, les limites de securite et les fenetres de contexte peuvent changer rapidement. Verifiez toujours les decisions de production avec le fournisseur avant d acheter, deployer ou migrer.
Sources officielles et techniques utiles
Current OpenAI API model families, context windows, prices and specialized models.
Gemini, Nano Banana, Veo and live/audio model listings.
Claude model comparison, context, max output and cloud availability.
Grok 4.3, Grok Build, Grok Imagine and xAI model guidance.
Mistral, Ministral, Devstral, Codestral, Voxtral and OCR models.
DeepSeek V4 Flash and V4 Pro context, output limits, features and pricing.