Home / AI Answers / AI Glossary / Large Language Models

Large Language Models

What is a Large Language Model

A Large Language Model (LLM) is an advanced type of Artificial Intelligence model that specializes in understanding, generating, and manipulating human languages. LLMs are typically based on deep learning architectures, such as Transformers and Recurrent Neural Networks (RNN), and are trained on massive datasets containing billions of words or tokens. Some popular LLMs include OpenAI’s GPT (Generative Pre-trained Transformer) series known as ChatGPT, Google’s Gemini, Anthropic’s Claude Meta’s Llama and Grok by xAI.

ELI5 Large Language Models Explained Like You’re Five

Imagine you have a talking parrot that has read a ton of books, listened to lots of stories, and learned a lot of words. Because of this, the parrot can understand what you say and even help you with your homework by giving you answers or telling you stories.

Large Language Models (LLMs) are like that super-smart parrot. They are computer programs that have read and learned from a huge amount of text, like books, websites, and articles. This helps them understand language and generate text that makes sense. They can help you write essays, answer questions, translate languages, and chat with you.

So, a Large Language Model is a super-smart computer program that understands and uses language really well because it has learned from reading lots of text.

Components

Large Language Models consist of several components that work together to process and generate human language:

  1. Tokenization: Tokenization is the process of breaking down the input text into smaller units, such as words or subwords, which are referred to as tokens.
  2. Embeddings: Embeddings are mathematical representations that convert tokens into continuous vectors, allowing the model to capture semantic relationships between words.
  3. Encoder: The encoder is responsible for processing the input text and generating contextualized representations for each token. In Transformer-based models, the encoder uses self-attention mechanisms to capture the relationships between tokens.
  4. Decoder: The decoder takes the contextualized representations generated by the encoder and produces the output text. In generative models, such as GPT, the decoder generates text by predicting the next token in the sequence based on the previous tokens.
  5. Fine-tuning: After pre-training on large datasets, LLMs can be fine-tuned for specific tasks or domains using smaller, task-specific datasets.

Applications and Impact

Large Language Models have a wide range of applications across various industries, including:

  • Natural Language Understanding (NLU): LLMs have significantly improved the ability of AI systems to understand human language by capturing complex linguistic structures and semantic relationships.
  • Natural Language Generation (NLG): LLMs can generate coherent and contextually relevant text, enabling applications such as content creation, summarization, and paraphrasing.
  • Machine Translation: LLMs have led to significant advancements in machine translation by enabling more accurate and fluent translations between languages, taking into account context and meaning.
  • Question Answering: LLMs can be used to build question-answering systems that can provide accurate and contextually relevant answers to user queries.
  • Sentiment Analysis: LLMs can analyze the sentiment and emotion expressed in text, enabling applications such as customer feedback analysis and social media monitoring.

Challenges and Limitations

Despite the impressive capabilities of Large Language Models, several challenges and limitations remain:

  1. Computational Resources: Training LLMs requires massive computational resources, including powerful Graphics Processing Units (GPUs) and specialized hardware, leading to high costs and environmental concerns.
  2. Data Requirements: LLMs require large amounts of data for training, which can be challenging to obtain and maintain, especially for low-resource languages and domains.
  3. Bias and Fairness: LLMs can inadvertently learn and perpetuate biases present in the training data, leading to potential ethical concerns and unfair treatment of certain groups or individuals.
  4. Interpretability and Explainability: The complexity of LLMs makes it difficult to understand how they arrive at specific outputs or decisions, which can be a concern for applications where transparency is crucial.
  5. Robustness and Adversarial Attacks: LLMs can be sensitive to small input perturbations or adversarial examples, potentially leading to incorrect or misleading outputs.
  6. Generalization: While LLMs can perform well on tasks closely related to their training data, their performance may degrade when faced with tasks or domains that are significantly different.

Real-world examples

Large Language Models have been successfully applied to a variety of real-world applications, including:

  • Customer Support: LLMs are used in Chatbots and Conversational AI systems to handle customer support inquiries, providing accurate and contextually relevant responses to user queries. For more see AI chatbots, AI customer service.
  • Content Generation: LLMs can generate high-quality text for various purposes, such as creating news articles, product descriptions, or social media posts.
  • Summarization: LLMs can automatically generate summaries of long documents, articles, or reports, helping users quickly grasp the main ideas and points. For more see AI Summarize.
  • Semantic Search: LLMs can be used to build search engines that understand the meaning and context of user queries, providing more accurate and relevant search results. For more see AI search engines.
  • Text Classification: LLMs can classify documents, emails, or social media posts into categories based on their content, enabling applications such as spam filtering and content moderation.

Future Developments

As research and development in Large Language Models continue, several future developments can be anticipated:

  1. Efficient Training and Inference: New techniques and algorithms will be developed to reduce the computational resources and energy required to train and deploy LLMs, making them more accessible and environmentally friendly.
  2. Improved Multilingual and Multimodal Capabilities: Future LLMs will likely be able to understand and process multiple languages and modalities (e.g., text, speech, images) simultaneously, further enhancing their capabilities and applications.
  3. Transfer Learning and Few-shot Learning: Advances in transfer learning and few-shot learning will enable LLMs to adapt to new tasks and domains with minimal additional training data, improving their generalization capabilities.
  4. Addressing Bias and Fairness: Researchers and developers will continue to develop methods to identify, quantify, and mitigate biases in LLMs, ensuring that they are used responsibly and fairly.
  5. Interpretability and Explainability: Techniques for enhancing the interpretability and explainability of LLMs will be developed, allowing users to better understand and trust their outputs and decisions.

In conclusion, Large Language Models have significantly advanced the capabilities of AI systems in understanding and generating human language, enabling a wide range of applications across various industries. As research and development in this area continue, LLMs will become increasingly sophisticated and capable, further unlocking the potential of AI-driven solutions for language-related tasks.

References

  1. Brown, T. B., Mann, B., Ryder, N., Subbiah, M., Kaplan, J., Dhariwal, P., … & Amodei, D. (2020). Language models are few-shot learners. arXiv preprint arXiv:2005.14165. Link
  2. Devlin, J., Chang, M. W., Lee, K., & Toutanova, K. (2018). BERT: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805. Link
  3. Raffel, C., Shazeer, N., Roberts, A., Lee, K., Narang, S., Matena, M., … & Liu, P. J. (2019). Exploring the limits of transfer learning with a unified text-to-text transformer. arXiv preprint arXiv:1910.10683. Link
  4. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., … & Polosukhin, I. (2017). Attention is all you need. In Advances in neural information processing systems (pp. 5998-6008). Link

LLM FAQ

Q: What are large language models in AI?

A: Large language models in AI are advanced deep learning models that specialize in understanding, generating, and manipulating human languages. They are trained on massive datasets and often based on architectures like Transformers or Recurrent Neural Networks.

Q: What is the purpose of large language models?

A: The purpose of large language models is to process, understand, and generate human languages for various applications, such as natural language understanding, natural language generation, machine translation, question answering, and sentiment analysis.

Q: What are the 4 types of AI models?

A: The four types of AI models are rule-based systems, machine learning models, deep learning models (including neural networks), and hybrid systems that combine multiple approaches.

Q: What is the difference between a large language model and a neural network?

A: A large language model is a specific type of deep learning model that focuses on processing human languages, while a neural network is a more general term that refers to a class of algorithms used in a wide range of AI applications, including large language models.

Q: What is the power of large language models?

A: The power of large language models lies in their ability to capture complex linguistic structures and semantic relationships, enabling them to understand and generate human languages more accurately and contextually relevant than previous AI models.

Q: How do large language models write code?

A: Large language models write code by predicting the most likely sequence of tokens (words or symbols) based on the input context and their knowledge of programming languages. They use their deep understanding of syntax, semantics, and common programming patterns to generate code that is coherent and contextually relevant.

Q: How accurate are large language models?

A: The accuracy of large language models depends on the specific task, domain, and quality of the training data. In general, LLMs have shown impressive performance in various language-related tasks, such as natural language understanding, machine translation, and question answering. However, their accuracy may degrade when faced with tasks or domains significantly different from their training data or when encountering adversarial examples and input perturbations.

List of Large Language Models

  1. BERT (2018, Google): 340 million parameters, trained on 3.3 billion words. An early and influential language model.
  2. XLNet (2019, Google): Approximately 340 million parameters, trained on 33 billion words. An alternative to BERT.
  3. GPT-2 (2019, OpenAI): 1.5 billion parameters, trained on ~10 billion tokens. General-purpose model based on transformer architecture.
  4. GPT-3 (2020, OpenAI): 175 billion parameters, trained on 300 billion tokens. A fine-tuned variant, GPT-3.5, was made public through ChatGPT in 2022.
  5. GPT-Neo (March 2021, EleutherAI): 2.7 billion parameters. The first of a series of free GPT-3 alternatives.
  6. GPT-J (June 2021, EleutherAI): 6 billion parameters. A GPT-3-style language model.
  7. Megatron-Turing NLG (October 2021, Microsoft and Nvidia): 530 billion parameters, trained on 338.6 billion tokens. Standard architecture trained on a supercomputing cluster.
  8. Ernie 3.0 Titan (December 2021, Baidu): 260 billion parameters, trained on 4 Tb. Chinese-language LLM.
  9. Claude (December 2021, Anthropic): 52 billion parameters, trained on 400 billion tokens. Fine-tuned for desirable behavior in conversations. (see Claude)
  10. GLaM (December 2021, Google): 1.2 trillion parameters, trained on 1.6 trillion tokens. Sparse mixture-of-experts model.
  11. Gopher (December 2021, DeepMind): 280 billion parameters, trained on 300 billion tokens.
  12. LaMDA (January 2022, Google): 137 billion parameters, trained on 1.56T words, 168 billion tokens. Specialized for response generation in conversations.
  13. GPT-NeoX (February 2022, EleutherAI): 20 billion parameters. Based on the Megatron architecture.
  14. Chinchilla (March 2022, DeepMind): 70 billion parameters, trained on 1.4 trillion tokens. Used in the Sparrow bot.
  15. PaLM (April 2022, Google): 540 billion parameters, trained on 768 billion tokens. Aimed to reach the practical limits of model scale.
  16. OPT (May 2022, Meta): 175 billion parameters, trained on 180 billion tokens. GPT-3 architecture with adaptations from Megatron.
  17. YaLM 100B (June 2022, Yandex): 100 billion parameters, trained on 1.7TB. English-Russian model based on Microsoft’s Megatron-LM.
  18. Minerva (June 2022, Google): 540 billion parameters. LLM trained for solving “mathematical and scientific questions using step-by-step reasoning”.
  19. BLOOM (July 2022, Large collaboration led by Hugging Face): 175 billion parameters, trained on 350 billion tokens. Trained on a multi-lingual corpus.
  20. Galactica (November 2022, Meta): 120 billion parameters, trained on 106 billion tokens. Trained on scientific text and modalities.
  21. AlexaTM (November 2022, Amazon): 20 billion parameters, trained on 1.3 trillion tokens. Utilizes a bidirectional sequence-to-sequence architecture.
  22. LLaMA (February 2023, Meta): 65 billion parameters, trained on 1.4 trillion tokens. Trained on a large 20-language corpus for better performance with fewer parameters.
  23. GPT-4 (March 2023, OpenAI): Exact number of parameters unknown, approximately 1 trillion. Available for ChatGPT Plus users and used in several products.
  24. Cerebras-GPT (March 2023, Cerebras): 13 billion parameters. Trained with the Chinchilla formula.
  25. Falcon (March 2023, Technology Innovation Institute): 40 billion parameters, trained on 1 Trillion tokens. The model uses significantly less training compute than several other models.
  26. BloombergGPT (March 2023, Bloomberg L.P.): 50 billion parameters, trained on a 363 billion token dataset from Bloomberg’s data sources, plus general purpose datasets. Specialized for financial data.
  27. PanGu-Σ (March 2023, Huawei): 1.085 trillion parameters, trained on 329 billion tokens.
  28. OpenAssistant (March 2023, LAION): 17 billion parameters, trained on 1.5 trillion tokens. Trained on crowdsourced open data.
  29. GPT-4 (March 2023, OpenAI): Exact number of parameters unknown, approximately 1 trillion. Available for ChatGPT Plus users and used in several products.
  30. Cerebras-GPT (March 2023, Cerebras): 13 billion parameters. Trained with the Chinchilla formula.
  31. Falcon (March 2023, Technology Innovation Institute): 40 billion parameters, trained on 1 Trillion tokens. The model uses significantly less training compute than several other models.
  32. BloombergGPT (March 2023, Bloomberg L.P.): 50 billion parameters, trained on a 363 billion token dataset from Bloomberg’s data sources, plus general purpose datasets. Specialized for financial data.
  33. PaLM 2 (May 2023, Google): Exact number of parameters unknown, trained on a larger and more diverse corpus than its predecessor, PaLM. Excels at advanced reasoning tasks, translation, and code generation. Demonstrates improved multilingual capabilities and a more efficient architecture thanks to the use of compute-optimal scaling and an improved dataset mixture. It powers generative AI features at Google, like Bard and the PaLM API.
  34. Claude 2 (June 2023, Anthropic): 70 billion parameters, trained on an extensive dataset allowing for a 100,000-token context window. This model excels in long-form document processing and conversational AI tasks. It features improved coding skills and safety measures to minimize harmful outputs.
  35. Pythia (June 2023, EleutherAI): Suite of LLMs of different sizes trained on public data to help researchers understand LLM training processes.
  36. MPT (June 2023, MosaicML): 7B and 30B parameter models trained on 1T tokens of English and code. Licensed for commercial use.
  37. Falcon 180B (July 2023, Technology Innovation Institute): 180 billion parameters, trained on 1 to 1.5T tokens of English and code.
  38. StableLM (August 2023, StabilityAI): 3B and 7B parameter models trained on 1.5T tokens of an experimental dataset built on ThePile, followed by a v2 series with a diverse data mix.
  39. X-Gen (July 2023, Salesforce): 7B parameter models trained on 1.5T tokens of natural language and code.
  40. LLaMA 2 (July 2023, Meta): 7 to 70B parameter models trained on 2T tokens from publicly available sources, with extensive finetuning from human-preferences.
  41. Mistral 7B (September 2023, Mistral): 7 billion parameters, trained on an undisclosed number of tokens from open web data (see Mistral).
  42. DeciLM (October 2023, Deci.AI): Large model with undisclosed parameters and data sources.
  43. Qwen (October 2023, Alibaba): Bilingual English-Chinese models with 7 to 70 billion parameters trained on 2.4T tokens (see Qwen).
  44. Yi (November 2023, 01-AI): Bilingual English-Chinese models with 6 to 34 billion parameters trained on 3T tokens.
  45. Grok-1 (November 2023, xAI): 314 billion parameters (see Grok).
  46. DeepSeek (December 2023, DeepSeek AI): Coding model trained from scratch on 2T tokens, primarily focused on code with some natural language.
  47. LLaMA 3 (November 2023, Meta): 8 billion, 70 billion, and a forthcoming 400 billion parameters model.
  48. Gemma (December 2023, Google): 2 billion and 7 billion parameters, trained for a variety of tasks including conversational AI.
  49. Claude 3 (December 2023, Anthropic): Parameters unknown. Designed to be safe and reliable for enterprise use.
  50. Stable LM 2 (January 2024, Stability AI): Open-source, 12 billion parameters.
  51. Gemini 1.5 (February 2024, Google): Introduced as a more powerful version of the initial Gemini models, featuring advancements like a mixture-of-experts approach and a large context window up to one million tokens. Two notable variants are Gemini 1.5 Pro and Gemini 1.5 Flash.
  52. Gemma (February 2024, Google): A family of open-source lightweight LLMs derived from the Gemini models, available in two sizes with 2 and 7 billion parameters respectively. These models are designed for broader accessibility and versatility.
  53. Mixtral 8x22B (April 10, 2024, Mistral AI): Open-source, 141 billion parameters.
  54. GPT-4o (May 2024, OpenAI): GPT-4o, where the “o” stands for “omni,” is a multimodal model capable of handling text, speech, and video. It provides GPT-4-level intelligence but is much faster and improves on its capabilities across text, voice, and vision.
  55. PHI-2 (2024, Microsoft): A 2.7 billion parameter model that leverages high-quality training data and innovative scaling methods to outperform larger models in various benchmarks. PHI-2 demonstrates that smaller, well-trained models can achieve competitive performance.
  56. Claude 3.5 (June 2024, Anthropic): unknown parameters.
  57. LLamA 3.1 (June 2024, Meta AI): 405 billion parameters.
  58. Grok-2 (August 2024, xAI): including FLUX.1
  59. OpenAI o1-preview (September 2024, OpenAI)
  60. Tencent Hunyuan-Large (November 2024, Tencent): 389B param MoE, open-sourced.
  61. DeepSeek V3 (December 2024, DeepSeek) Claims better performance than OpenAI’s o1 (see DeepSeek).
  62. Kimi k1.5 (January 2025, Moonshot AI): A multimodal reasoning LLM with 128k token context window (see Kimi AI).
  63. Qwen 2.5-Max (January 2025, Alibaba): Multimodal text/video/image.
  64. Grok-3 (February 2025, xAI)
  65. Gemma 3 (March 2025, Google): Open model family released in 1B, 4B, 12B, and 27B sizes.
  66. Command A (March 2025, Cohere): Enterprise-focused large language model.
  67. Gemini 2.5 Pro Experimental (March 2025, Google DeepMind): Experimental “thinking” model in the Gemini 2.5 line.
  68. Qwen2.5-VL-32B-Instruct (March 2025, Alibaba): Vision-language large model.
  69. Qwen2.5-Omni-7B (March 2025, Alibaba): Multimodal model supporting text, image, video, and audio.
  70. Llama 4 (April 2025, Meta): Multimodal Mixture-of-Experts model family.
  71. GPT-4.1 (April 2025, OpenAI): Updated GPT-4-series model family.
  72. OpenAI o3 (April 2025, OpenAI): Reasoning-focused large language model.
  73. OpenAI o4-mini (April 2025, OpenAI): Compact reasoning model with vision support.
  74. Qwen3 (April 2025, Alibaba): New open-weight Qwen model family.
  75. Claude Opus 4 (May 2025, Anthropic): High-end Claude 4 series model.
  76. Claude Sonnet 4 (May 2025, Anthropic): Mid-tier Claude 4 series model.
  77. Magistral Small (June 2025, Mistral): Open reasoning-oriented language model.
  78. Magistral Medium (June 2025, Mistral): Commercial reasoning-oriented language model.
  79. Baidu ERNIE 4.5 / X1 (June 2025, Baidu): ERNIE 4.5 open-sourced; X1 reasoning variant released alongside.
  80. Grok 4 (July 2025, xAI): Successor to Grok-3 with expanded reasoning and multimodal capabilities.
  81. Kimi K2 (July 2025, Moonshot AI): 1T param MoE — base K2 release, predates K2.5.
  82. gpt-oss-120B (August 2025, OpenAI): Open-weight large language model.
  83. gpt-oss-20B (August 2025, OpenAI): Smaller open-weight companion model.
  84. Claude Opus 4.1 (August 2025, Anthropic): Incremental update to Opus 4.
  85. GPT-5 (August 2025, OpenAI): Next-generation flagship multimodal language model.
  86. Apertus (September 2025, Swiss National AI Initiative): Fully open, multilingual European large language model.
  87. Qwen3-Max (September 2025, Alibaba): High-capacity Qwen3 variant.
  88. Qwen3-Next (September 2025, Alibaba): Next-generation Qwen3 architecture.
  89. Qwen3-Omni (September 2025, Alibaba): Fully multimodal Qwen3 model.
  90. Claude Haiku 4.5 (October 2025, Anthropic): Fastest, lowest-cost tier in the Claude 4 generation.
  91. IBM Granite 4.0 (October 2025, IBM): Hybrid Mamba/transformer architecture for enterprise efficiency.
  92. MiniMax M2 (October 2025, MiniMax): 230B params (10B active) MoE, agentic coding/tool use, MIT license.
  93. GPT-5.1 (November 2025, OpenAI): Iterative improvement over GPT-5.
  94. Gemini 3 Pro (November 2025, Google DeepMind): First major release in the Gemini 3 series.
  95. Claude Opus 4.5 (November 2025, Anthropic): Further upgraded Claude 4 model.
  96. Mistral 3 (December 2025, Mistral): Third-generation Mistral model family.
  97. GPT-5.2 (December 2025, OpenAI): Enhanced GPT-5 series release.
  98. GPT-5.2-Codex (December 2025, OpenAI): Code-specialized GPT-5.2 variant.
  99. NVIDIA Nemotron 3 Nano (December 2025, NVIDIA): 31.6B total parameters (3.2B active), hybrid Mamba-Transformer MoE, 1M token context.
  100. Amazon Nova 2 (December 2025, Amazon): Family of 4 — Nova 2 Lite, Nova 2 Pro (preview), Nova 2 Omni, Nova 2 Sonic. 1M token context, configurable extended thinking (low/medium/high).
  101. Tencent Hunyuan 2.0 (December 2025, Tencent): Latest flagship update, integrated across WeChat/Yuanbao.
  102. Kimi K2.5 (January 2026, Moonshot AI): 1 trillion parameters (32B active), 262K context. Native multimodal, agent swarm technology coordinating up to 100 parallel sub-agents.
  103. Baidu ERNIE 5.0 (January 2026, Baidu): 2.4T-param omnimodal flagship, native full-modal unified modeling.
  104. GLM-5 (February 2026, Zhipu AI/Z.AI): MoE architecture, MIT licensed. Leading open-weight model on agentic coding benchmarks at release.
  105. Qwen3.5 (February 2026, Alibaba): 397B total/17B active flagship (Plus variant), hybrid linear attention (Gated Delta Networks) with sparse MoE, 262K context extendable toward 1M, 200+ languages.
  106. Gemini 3.1 Pro (February 2026, Google DeepMind): Iteration on Gemini 3 Pro with 1M token context, 65,536 token output, introduces a “medium” thinking level.
  107. Claude Opus 4.6 (February 2026, Anthropic): Large single-generation gain on ARC-AGI-2 (58.3%, up from 13.6%).
  108. Claude Sonnet 4.6 (February 2026, Anthropic): Mid-tier Claude 4 series model matching Opus 4.5 on long-horizon coding.
  109. ByteDance Doubao-Seed 2.0 (February 2026, ByteDance): Pro/Lite/Mini/Code variants, 256K context, powers Doubao (155M weekly active users).
  110. GPT-5.3 (~February/March 2026, OpenAI): Interim GPT-5.x release; instant variant became more accurate and better contextualized for search.
  111. GPT-5.4 (March 2026, OpenAI): Built-in computer use, improved deep research, 33% reduction in factual errors vs GPT-5.2. Mini and nano variants followed.
  112. Mistral Small 4 (March 2026, Mistral): Merges Magistral (reasoning), Pixtral (vision), and Devstral (coding) into one model with configurable reasoning effort.
  113. Phi-4-reasoning-vision-15B (March 2026, Microsoft): 15B params, combines vision + reasoning, decides autonomously when to reason deeper.
  114. NVIDIA Nemotron 3 Super (March 2026, NVIDIA): 120B total parameters (12B active), optimized for collaborative multi-agent workloads.
  115. GLM-5.1 (April 2026, Zhipu AI/Z.AI): 744B parameter MoE, MIT licensed, supports 8+ hour autonomous agentic tasks.
  116. Qwen3.6 (April 2026, Alibaba): Preview with dense 27B and 35B-A3B MoE variants; improved agentic coding.
  117. Kimi K2.6 (April 2026, Moonshot AI): 1T parameter native multimodal vision-language model, first open-weight system to beat GPT-5.4 on SWE-Bench Pro.
  118. DeepSeek-V4-Pro / DeepSeek-V4-Flash (April 2026, DeepSeek): 1.6T/49B active and 284B/13B active MoE models, 1M token context, hybrid attention architecture, MIT licensed.
  119. Claude Opus 4.7 (April 2026, Anthropic): Iteration on Opus 4.6.
  120. GPT-5.5 (April 2026, OpenAI): Emphasis on long-horizon agentic tasks, computer use, and error recovery mid-task.
  121. Grok 4.3 (April 2026, xAI): 1M token context, native video input, generates files (PDFs, slides, spreadsheets) directly.
  122. IBM Granite 4.1 (April 2026, IBM): Family of 10 models — 3B/8B/30B LLMs (full-precision + FP8), safety model, vision-language model, speech model. 512K context, ~15T training tokens.
  123. Qwen3.7-Max (May 2026, Alibaba): Closed-weight reasoning agent, 1M token context.
  124. Claude Opus 4.8 (May 2026, Anthropic): Improved honesty/calibration, dynamic workflows (parallel subagents) in Claude Code, fast mode.
  125. Baidu ERNIE 5.1 (May 2026, Baidu): Parameter-efficient follow-up to 5.0 — ~1/3 total params, ~1/2 active params, ~6% of comparable pretraining cost.
  126. Claude Fable 5 (June 2026, Anthropic): First Mythos-class model, generally available; always-on adaptive thinking, 1M context, 128K output, safety classifiers fall back to Opus 4.8 for flagged domains.
  127. Claude Mythos 5 (June 2026, Anthropic): Same capability as Fable 5 without the safety classifiers; limited availability via Project Glasswing.
  128. Kimi K2.7 Code (June 2026, Moonshot AI): Agentic coding-focused variant of K2.
  129. GLM-5.2 (June 2026, Zhipu AI/Z.AI): Highest reported SWE-bench Pro among Chinese models at release.
  130. Qwen3.5-9B (June 2026, Alibaba): Lightweight native multimodal dense model, hybrid Gated Delta Networks/Gated Attention architecture.
  131. NVIDIA Nemotron 3 Ultra (June 2026, NVIDIA): 550B total parameters (55B active), Hybrid Mamba-Attention with LatentMoE routing, 1M token context.
  132. Claude Sonnet 5 (June 2026, Anthropic): Latest Sonnet-tier release.

Curious about diving deeper into the world of artificial intelligence?

Discover key terms and concepts that shape the AI landscape.
AI Glossary

Your trusted SFW AI Directory.

Navigate

About us

FAQ

Contact Us

Affiliates

explore

Most Popular AI Tools

AI Tools Hall of Fame

AI Tools by AI Model

AI Tools by Integration

AI Tools by Compliance

Free AI Tools

List your tool

Submit AI Tool

© 2023-2026 AI Tools Explorer. All rights reserved. By accessing this site, you agree to our Terms of Service and Privacy Policy. AI Tools Explorer may receive a share of revenue from items bought via our site, as part of our Affiliate Partnerships with retailers. Unauthorized reproduction, distribution, transmission, caching, or any other usage of the content on this site is strictly prohibited, unless expressly permitted in writing by AI Tools Explorer.

Privacy Policy | Terms of Service | Do Not Sell or Share My Personal Information | Affiliate disclosure