A list of large language models (LLMs)

As the technology matures, new LLMs continue to proliferate. Leading AI developers, new start-ups and established enterprise powerhouses alike are perpetually releasing and refining new models. Meanwhile, the open source community is constantly fine-tuning open source LLMs, merging and modifying existing models on custom datasets to create endless variants. As such, no list of LLMs could reasonably hope to be exhaustive—and even the most “exhaustive” list wouldn’t remain so for very long.

What follows is a list of some of the most prominent and performant LLMs available today. Here are some things to note:

• This list is limited to generative LLMs capable of text generation, excluding AI models that are designed only for natural language understanding (NLU) or for other natural language processing (NLP) tasks. It does not include, for instance, the embedding models (such as BERT or Granite Embedding) that power semantic search and retrieval augmented generation (RAG).

• The list prioritizes models that are actively being supported and updated by their developers, and maintain at least nominally competitive performance. This excludes a number of historically influential foundation models, such as Google’s T5, OpenAI’s GPT-3 or Meta’s Llama 2, some of which continue to be used for research purposes.

For practical purposes, LLMs can generally be divided into 2 categories: closed source LLMs, available solely as commercial offerings through the model developer, and open models, which are made freely available at no cost.

A closed source model, or proprietary model, can only be accessed directly on the model developer’s platform, other platforms to which they have licensed their model or through the model provider’s proprietary API.

Because closed model developers generally treat their technical details as closely guarded trade secrets, it’s typically impossible to know with certainty the specifics of a closed model’s size, neural network architecture or training process. Some details can be inferred—for instance, by comparing a closed model’s inference speed, GPU memory usage and benchmark performance to that of open models whose details are publicly disclosed—but rarely, if ever, confirmed.

Since at least since roughly 2022, most state-of-the-art frontier models at any given time have been closed models—but that’s largely a reflection of the real-world historical circumstances of the industry, rather than any inherent superiority of closed models to open models. What follows are some of the most notable closed model series, ordered alphabetically.

Anthropic’s Claude language models are among the world’s most performant. Originally founded as an AI safety research lab in 2021 by former OpenAI employees, Anthropic’s approach to model development is built around the unique concept of Constitutional AI. Claude’s “Constitution” is a document that serves not only to guide the conduct of Anthropic employees, but the conduct (and creation of synthetic training data) of Claude models themselves.

Since Claude 3, successive generations of Claude have featured multimodal models in 3 different sizes:

• Claude Haiku are Anthropic’s smallest models, optimized for speed and cost-efficiency. Unlike Sonnet and Opus, Haiku models are not reasoning models: unless explicitly prompted to do so, Haiku models do not output chain-of-thought (CoT) reasoning traces.

• Claude Sonnet are Anthropic’s mid-sized models, aimed at what Anthropic considers to be the optimal tradeoff between performance and efficiency for most use cases. Both Sonnet and Opus are hybrid reasoning models, meaning they can be configured to perform either standard inference or adaptive CoT reasoning for complex, multi-step problem solving.

• Claude Opus are Anthropic’s largest, most powerful models, aimed at frontier performance across challenging tasks.

Claude Haiku, Sonnet and Opus can all process text, audio and image inputs, and output text or audio (as text-to-speech). Historically, unlike most of their closed model competitors, they (and the Claude platform that they power) were not capable of image generation—but as of March 12, 2026, Claude can now generate images. When accessing the models through the Claude API, users can set the “effort level” of Sonnet or Opus’s reasoning process to “max,” “high,” “medium,” “low” or “adaptive.”

Gemini is Google’s closed language model series, developed by its subsidiary Google DeepMind and first launched in December 2023. It’s worth noting that Google Brain (which was merged with DeepMind to form Google DeepMind in 2023) is responsible for the creation of the transformer model architecture that enabled the first LLMs, having published the landmark “Attention is All You Need” research paper in 2017.

Since early 2025, Google has released each generation of Gemini models with 3 different sizes, all of which are reasoning models. When accessed through the Gemini API, users can select one of multiple “thinking levels” to customize the amount of tokens and time the model will spend before generating a final output.

• Gemini Pro models are Google’s largest, state-of-the-art LLMs.
  
  
• Gemini Flash models are, in comparison to Gemini Pro, optimized for speed.
  
  
• Gemini Flash-Lite models are fast, cost-efficient models optimized for high-volume tasks such as translation and agentic tool use.

Gemini Pro, Flash and Flash-Lite models are natively multimodal: they can process text, audio, image or video inputs and generate text outputs. When accessed through the Gemini platform, multimodal outputs can be generated through Gemini’s separate, specialized models for image generation, video generation or music generation.

Since the release of Gemini 2.5 Pro in March 2025, which achieved then-industry best performance across most academic benchmarks, Gemini models have contended with Claude and OpenAI’s GPT series as the world’s most performant LLMs. Generally speaking, the status of “top” model changes hands each time a new frontier model in one those three series is released.

Grok is a family of proprietary LLMs produced by xAI, first launched in beta preview as a chatbot on X (formerly Twitter) in November 2023. In April 2025, xAI launched API access for Grok 3, which was then its newest, flagship model.

Grok’s model lineup has continued to change over successive generations of model releases.

• Grok 2 was accompanied by Grok 2 Mini, the model family’s first size-based variant. This same convention was repeated for Grok 3 in February 2025.
  
  
• The 4th generation of Grok models was launched with Grok 4 and Grok 4 Heavy in July 2025. In the fall of 2025, these were followed by Grok 4 Fast, and then by Grok 4.1 (available in both Thinking and Non-thinking configurations).
  
  
• In August 2025, xAI released Grok Code Fast 1, an efficiency-focused model optimized for agentic coding.

As of Grok 4, Grok models can process text, image and speech inputs. Though the Grok LLMs cannot provide multimodal outputs, image and video outputs can be generated by xAI’s Aurora model through its Grok Imagine platform.

Unrelated to its raw performance, much of Grok’s history (and particularly that of the Grok chatbot) has been marked by controversy, such as accusations of spreading election misinformation, inserting polarizing viewpoints into unrelated conversations and perpetuating harmful stereotypes.

In public statements, xAI CEO Elon Musk has said that “our general approach is that we will open source the last version when the next version is fully out.”¹

xAI open-sourced Grok 1 under Apache 2.0 license in March 2024. Though Grok 3 was released in February 2025, the next open-source release of a Grok model was not until August of 2025. Confusingly, xAI (and Musk) announced that they had open-sourced “Grok 2.5,” ² though no model had been named nor announced as such prior to that statement. The model’s own Hugging Face model card even refers to the model as “Grok-2.”

In that August 2025 announcement, Musk indicated that Grok 3 would be likewise open-sourced in “about 6 months.” As of 8 months later, said open source release date is yet to be announced.

OpenAI’s GPT series—short for Generative Pretrained Transformer—is largely credited with initiating the current era of generative AI, particularly following the 2022 launch of ChatGPT with their GPT-3.5 model.

OpenAI’s for model naming and variant conventions has changed significantly since 2022, often in a confusing way. For instance, GPT-4.1 was released after GPT-4.5, and the o4 reasoning model was available at the same time as the multimodal non-reasoning model GPT-4o, which was an entirely distinct from their “o4” reasoning model, whose performance was inferior to that of “o3.” In early 2025, OpenAI CEO Sam Altman acknowledged that “We realize how complicated our model and product offerings have gotten.”

Since the release of GPT-5 in August 2025, the company’s consolidated LLM offerings now comprise:

• GPT-5.x is OpenAI’s flagship general-purpose offering. As of March 2026, the latest model version is GPT-5.4. Though all GPT-5 models are reasoning models, GPT-5.4 is also available in a GPT-5.4 Pro variant, which “uses more compute to think harder and provide consistently better answers.”³ OpenAI also offers GPT-5 Codex, a version of GPT-5 fine-tuned for optimal agentic code generation (which is periodically updated following updated versions of the core model).

• GPT-5 mini offers “near-frontier intelligence for cost sensitive, low latency, high volume workloads,” according to OpenAI’s model overviews.

• GPT-5 nano is the “fastest, most cost-effective version of GPT-5.”

OpenAI has also release 2 open weight GPT models, which are detailed in the “Open models” section of this article.

Mistral AI, a France-based company founded by former employees of Meta AI and Google DeepMind, was originally dedicated entirely to open source models upon the launch of its first model (Mistral 7B) in September 2023. Since then, Mistral has transitioned to a mixed model in which many of its offerings have open releases but select frontier models remain closed source.

As of March 2026, Mistral AI’s flagship proprietary LLMs include:

• Mistral Medium 3.1, a general-purpose multimodal model released in August 2025.
  
  
• Codestral, a coding-focused model “built specifically for high-precision fill-in-the-middle (FIM) completion.”⁴
  
  
• Magistral Medium 1.2, a reasoning model companion to Mistral Medium.

Mistral’s open weight model offerings are detailed later in this article.

In machine learning, open source is often used colloquially to refer to AI tools whose source code is made available free of charge, but the term is actually a formal designation stewarded by the Open Source Initiative (OSI). The OSI only certifies a given software license as “Open Source Initiative approved” if it deems said license to meet the ten requirements listed in the official Open Source Definition (OSD).

Most “open source” models do not meet all of those requirements. That being the case, the term open model (or open weight model) more accurately refers to any freely distributed LLM. Within the spectrum of open models is a great deal of variability. An open weight (but not open source) model can be used to run inference and can even be fine-tuned—but if its full source code is not provided, it can’t be modified beyond changes to the values of its weights through fine-tuning. Its license might prohibit the model’s use in some scenarios (such as commercial settings) or place other specific stipulations on its application.

A true open source model released with training code and a description of its training procedures, conversely, can be fully modified in any way and used without restriction. The most common and standardized open source licenses are the Apache 2.0 license and MIT license. It’s worth noting, however, that unless an open source model’s developer provides the details of its training data, the model itself is not fully reproducible.

Open source releases are integral to the continued development and improvement of LLMs, and are largely responsible for enabling their invention in the first place. Open models can typically be accessed through their model developer or through popular open source platforms such as GitHub or Hugging Face. What follows is a list of notable open model series, organized alphabetically.

Cohere, a Canada-based company whose founders include one of the authors of “Attention is All You Need,” was launched in 2019. Though the company releases detailed technical reports for each LLM and ostensibly releases them as open weight models, Cohere licenses their open releases under a modified version of the Creative Commons 4.0 license  that prohibits commercial use.

 Command is Cohere’s flagship foundation model series, designed for enterprise use cases.

• Command R was the first generation of Cohere’s enterprise models, launched in March 2024 with a 35 billion parameter model with an emphasis on RAG and tool use. Its release was soon followed by Command R+, a 104B variant, the following month. They were joined by the smaller Command R7B that December.
  
  
• Command A, the second generation of Cohere’s enterprise models, was released in March 2025 with a focus on business, STEM and coding tasks. The original 111B model was eventually released in variants including Command A Reasoning, Command A Translate (fine-tuned to optimize translation performance across 23 languages) and Command A Vision, a vision language model (VLM) that paired the LLM with a vision encoder.

In a March 2026 Reddit comment, Cohere CEO Aidan Gomez indicated that the company was actively developing the next generation of Command, and that they would be the organization’s first mixture of experts (MoE) models.

Aya is Cohere’s multilingual-focused model series, first launched in February 2024 with Aya 101—which, as its name suggests, was “capable of following instructions in 101 languages.”

• Aya Vision is a multimodal, multilingual VLM, offered in 8B and 32B variants, with capabilities across 23 different languages.
  
  
• Tiny Aya, released in February 2026 is a series of lightweight multilingual models with 3.35B parameters. Tiny Aya-Base is a pretrained model supporting over 70 languages. TinyAya-Global is its instruction-tuned counterpart, supporting 67 languages.
  
  
• The Tiny Aya release also contained specialized regional variants. TinyAya-Earth is optimized for African and West Asian languages; TinyAya-Fire is optimized for South Asian languages; TinyAya-Water is optimized for Asia-Pacific and European languages.

DeepSeek is an integral player in the open source ecosystem, contributing a number of innovations to LLM architectures and training processes. At times, its models’ performance have rivaled that of top closed models. Their LLMs—both weights and code—are open sourced under a standard MIT license. DeepSeek also frequently releases technical papers detailing their findings and techniques.

• DeepSeek-V3 is a large MoE model, with 671B total parameters (and 37B active parameters during inference), first released in late 2024. The model is often credited with bringing the mixture of experts architecture back into mainstream prominence.
  
  
• DeepSeek-R1 is a reasoning model, built by fine-tuning DeepSeek-V3 using then-novel reinforcement learning techniques. DeepSeek-R1 was a landmark in the history of open source LLMs. It not only rivaled the performance of OpenAI previously-unmatched o1 model, but provided technical paper with full details of DeepSeek’s training methodology. Its release directly inspired the first generation of open reasoning models.
  
  
• DeepSeek-V3.1, released in August 2025, is a hybrid reasoning model, configuring to run both standard inference and CoT reasoning. In essence, it combined DeepSeek-V3 and DeepSeek-R1 into a single model. It was most recently updated as DeepSeek-V3.2 in October 2025. Both DeepSeek-V3.1 and DeepSeek-V3.2 retain the 671B-37B MoE architecture of the original model.
  
  
• DeepSeek also released several “DeepSeek-R1-Distill” models, created by fine-tuning smaller Qwen and Llama models to emulate DeepSeek-R1 through knowledge distillation.

Despite periodic rumors of an impending DeepSeek-V4 (or “DeepSeek-R2”), their releases are yet to materialize.

The Falcon series of LLMs are developed by the UAE’s Technology Innovation Institute (TTI). Though TII’s first generational of models in 2023 was perhaps most notable for Falcon-180B, which at the time was one of the largest open source models available, TII have since focused on smaller models. Falcon2 had 11B parameters and Falcon3, TII’s first multimodal models (released in December 2024), ranged from 1B to 10B.

The most recent generations of Falcon models have focused on hybrid Mamba-Transformer models.

• Falcon-H1, released in May 2025, comprises pretrained and instruction-tuned hybrid models in 0.5B, 1.5B, 3B, 7B and 34B variants. Falcon-H1R, released in January 2026, is a reasoning variant of Falcon-H1-7B.
  
  
• Falcon-H1-Tiny are, as their name suggests, extremely small variants of Falcon-H1, in sizes of 90M, 100M and 0.6B parameters. Each size is offered as both base modles and as specialized variants fine-tuned for specific use cases.
  
  
• Falcon-Edge models is a family of experimental 1-bit (“BitNet”) LLMs in sizes of 1B and 3B.

Falcon models are released under a proprietary Falcon license that is inspired by, but adds notable stipulations and constraints to, the Apache 2.0 framework.

Gemma is Google’s family of open models. According to Google, Gemma models are “built from the same technology that powers [their] Gemini models.”⁷

• Gemma 3, released in March 2025, is the latest generation of Gemma’s core LLM. The initial Gemma release contained both pretrained and instruction-tuned variants in sizes of 1B, 4B, 12B and 27B. In August 2025, Google added a smaller variant with 270M parameters. Gemma 3 models can process text or image inputs and offer multilingual support for over 140 languages.

• Gemma 3n, released in July 2025, feature an experimental MatFormer architecture that essentially allows for any number of smaller, custom-sized models to be “nested” within a single, larger model. The architecture is named after Russian nesting dolls, also called “Matryoshka” dolls—hence MatFormer. Gemma 3n is offered in nominal sizes of 2B and 4B parameters and supports text, image, video or audio inputs (but text-only outputs).
  
  
• FunctionGemma is a variant of Gemma 3 270M fine-tuned for tool use (or “function calling,” hence the name).

Gemma models are released under the Gemma license, whose usage terms are similar to those of the Apache 2.0 license but are governed by the Gemma Prohibited Use Policy.

GLM is a family of LLMs from Beijing-based Z.ai (also called Zhipu AI) that aim for state-of-the-art performance. The company achieved a breakthrough with GLM-4.5, which upon its initial release in late July 2025 ostensibly rivaled the world’s top open models, including the flagship models from DeepSeek and Qwen, across academic benchmarks.

• GLM-4.5 was offered in 2 model sizes—the flagship LLM, a large-scale MoE model with 355B total parameters (32B active), and the smaller GLM-4.5-Air (with 106B total parameters, 12B active). GLM-4.5V is a VLM, built on the GLM-4.5-Air foundation model, that adds computer vision and video understanding capabilities.
  
  
• GLM-4.6, an updated version of GLM-4.5 released on September 30, 2025, did not include a smaller text-only variant. However, in early December the company released GLM-4.6V (an update of GLM-4.5V) and GLM-4.6V-Flash, a 9B dense model.
  
  
• GLM-4.7, an update to the flagship text-only model released in late December 2025, added GLM-4.7-Flash, a significantly smaller LLM with only 30B total parameters (and 3B active parameters).
  
  
• GLM-5, released in February 2026, is significantly larger than its predecessors, with 744B total parameters (40B) active.

IBM Granite is a series of open source LLMs optimized for enterprise use cases, focused primarily on small, practical and efficient models. First launched in September 2023, Granite rose to prominence upon the release of Granite 3.0 in October 2024, which saw the Granite series reach performance rivaling that of leading open models of comparable size.

Granite 4, launched in October 2025, introduced a new hybrid Mamba2-Transformer architecture for superior speed and memory efficiency, particularly under large workloads, compared to conventional transformer models.

• Granite 4-H Small is a hybrid MoE model with 32B total parameters (9B active). Granite 4 also includes another hybrid MoE, Granite 4-H Tiny, with 7B total parameters (1B active), and a dense hybrid model, Granite 4-H Micro, with 3B active parameters.
  
  
• Granite 4 Micro is a 3B dense model built on a conventional transformer model architecture, unlike the 4-H models.
  
  
• Granite 4 Nano is a series of hybrid Mamba-transformer and conventional transformer models in sizes ranging from 350M parameters to 1B parameters.
  
  
• Granite 4 1B-Speech is a speech-to-text model designed for automatic speech recognition (ASR) and bidirectional automatic speech translation (AST).

All Granite models are open sourced under a standard Apache 2.0 license and trained on enterprise-safe data. In October 2025, the Granite series became the first major open model family to receive ISO-42001 certification.

GPT-OSS are OpenAI’s open weight language models, released in August 2025 under a standard Apache 2.0 license. They’re the company’s first open LLMs since the release of GPT-2 in 2019.

• GPT-OSS-120B is an MoE model 117B total parameters (5.1B active), designed for general purpose use and tasks benefitting from high-level reasoning.
  
  
• GPT-OSS-20B is an MoE model with 21B parameters (3.6B active) intended for lower latency use and local deployment.

Both GPT-OSS models were trained with 4-bit quantization of their model weights, significantly increasing their speed and reducing their memory requirements relative to those of conventional models of similar size.

Kimi is a series of open models developed by Beijing-based Moonshot AI.

• Kimi-K2 is a text-only, massive MoE model with 1 trillion total parameters (32B active). It achieved significant notoriety upon its release in July 2025 for rivaling and (sometimes beating) GPT-4.1 and Claude Opus 4 models on key coding benchmarks.⁸ 
  
  
• Kimi-K2 Thinking, the reasoning model variant of Kimi-K2, likewise caused a stir for once again rivaling top closed models across challenging agentic AI benchmarks.⁹ 
  
  
• Kimi-K2.5 an update to Kimi-K2 that adds multimodal vision capabilities. It can be operated in multiple “modes,” each optimized for specific use cases.

Kimi models are released under a modified MIT License, requiring users to “prominently display ‘Kimi K2’ on the user interface” of any product with over 100 million monthly active users or more than USD 20M in monthly revenue.

Meta’s Llama models (original stylized as LLaMA, short for “Large Language model Meta AI), have been an integral part of the history of open LLMs. Early Llama releases help democratize LLM methodologies, informing and strongly influencing many standard conventions of LLM development, from training to architecture and sizing variations.

• Llama 2, released in July 2023, was released in sizes of 7B, 13B and 70B.
  
  
• Llama 3, released in April 2024 in sizes of 8B and 70B, competed with many leading closed models across academic benchmarks. Llama 3.1 significantly expanded the models’ context length and added a then-unprecedentedly large 405B variant that July. Llama 3.2 added both smaller variants and vision capabilities, while Llama 3.3 featured a single 70B model whose performance rivaled that of Llama 3.1 405B.
  
  
• Llama 4, released in featured 2 large multimodal MoE models: Llama 4 Maverick, with 400B total parameters (17B active) and Llama 4 Scout, with 109B total parameters (19B active). Though their performance significantly exceeded that of prior Llama generations across most benchmarks, Llama 3 models remain Meta’s most popular LLMs (as reflected by downloads on Hugging Face).¹⁰

Though Meta often uses the term “open source,” Llama models are released under a custom Llama license that places constraints on usage, attribution and access. The Open Source Initiative has therefore criticized Meta’s use of the term.

Shanghai-based MiniMax Group released their first eponymous LLM, MiniMax-Text-01, and a companion VLM, MiniMax-VL-01, in January 2025. They have since risen to prominence as one of the premier LLM developers in China, prioritizing large-scale models and long context windows.

• MiniMax-M1, released in June 2025, is a text-only reasoning model built from fine-tuning MiniMax-Text-01. Like its predessor, it’s a large MoE model with 456B total parameters and 45.9B parameters activated per token.

• MiniMax-M2 offers superior performance and efficiency compared to M1. It has 230B total parameters, and a more fine-grained MoE architecture that activates only 10B parameters per token. Released in October 2025, it was updated as MiniMax-M2.1 two months later. MiniMax also offers MiniMax-M2-her, a version fine-tuned for character-based roleplay.
  
  
• MiniMax-M2.5 and MiniMax-M2.5-Lightning, released in February 2026, achieve further performance optimization, rivaling Claude Opus 4.5 on select coding benchmarks. They are identical in all regards except speed and throughput: the “Lightning” variant generates output twice as fast.
  
  
• MiniMax-M2.7, released in March 2026, is an update to MiniMax-M2.5 that the company claims helped to train itself.¹¹

MiniMax models are offered under a modified MIT License.

Alongside its closed-source offerings, Mistral AI offers a variety of well-regarded open models. Most (but not all) of Mistral’s open models are released under standard Apache 2.0 license.

• Mistral Large 3 utilizes a DeepSeek-V3-inspired MoE architecture, with 675B total parameters (41B active). Its benchmark performance is roughly equivalent to that of DeepSeek-V3.1 and Kimi-K2.1.¹² Released in December 2025, it’s multilinigual and multimodal, capable of processing both text and image inputs.
  
  
• Ministral 3 is Mistral’s small model series, offered in 3B, 8B and 14B sizes and base, instruction-tuned and reasoning variants.
  
  
• Mistral Small 3.2 is a 24B LLM released in June 2025. Its performance is comparable to that of the more recent Ministral 3 14B.
  
  
• Devstral is Mistral’s agentic engineering-focused model series. Devstral 2, released in December 2025, comprises two models. Devstral 2 123B is released under a modified MIT License, requiring organizations with over $20M USD in monthly revenue to request a commercial license from Mistral. Devstral Small 2 24B is released under standard Apache 2.0 license.
  
  
• Mixtral, released in December 2023, is an LLM that originally popularized the mixture of experts architecture for language models. As of early 2026, its 8x7B variant remains extremely popular on Hugging Face, with over 700,000 monthly downloads.¹³

Preeminent hardware manufacturer NVIDIA’s open LLM series are well regarded for their performance, research literature and architectural innovations.

• NVIDIA-Nemotron-Nano v2 is a family of hybrid Mamba-2-LLM models in sizes of 9B and 12B, capable of both reasoning and standard inference. They were released in August 2025 under a custom NVIDIA Open Model License Agreement with notable conditions regarding legal liabilities, usage and NVIDIA’s right to make future modifications to the agreement.
  
  
• Nemotron 3 Nano, released in Decemember 2025, comprises 2 models: Nemotron-3-Nano-4B and Nemotron-3-Nano-30B-A3B, an MoE with 30B total parameters (3B active). They were released under the NVIDIA Nemotron Open Model License, which omits NVIDIA’s right to make unilateral future updates to the terms.
  
  
• Nemotron 3 Super is a larger MoE with 120B total parameters (12B active), released in March 2026.

Olmo, developed by the Allen Institute for AI (“Ai2”), are among the most truly “open” of all open source models: Ai2 typically releases all code, weights, training checkpoints and associated datasets alongside a standard Apache 2.0 release.

• Olmo 3, released in November 2025, comprises dense transformer models in sizes of 7B and 32B. The models are released in base, instruct and “think” variants. In December 2025, the 32B received an update as Olmo 3.1.

• Olmo Hybrid, released in March 2026, is a 7B model with an experimental hybrid architecture combining both transformer and linear RNNs (based on the Gated DeltaNet architecture popularized by Qwen).

Phi is Microsoft’s open model line, historically focused on small models. They’re released under standard MIT License.

• Phi 4 is a 14B text-only LLM, originally released in December 2024.
  
  
• Phi 4-mini, released in in February 2025, is a smaller, 3.8B model.
  
  
• Phi 4-multimodal, released alongisde Phi 4-mini, supports text, image and speech inputs.
  
  
• Phi 4-Reasoning-Vision, released in March 2026 is a 15B model that adds holistic, multimodal reasoning across images, text and documents.

The Qwen series of LLMs, developed by Alibaba, have become among the most popular open models in the industry. The model family offers a wide variety of model sizes, architecture and capabilities intended to suit a variety of developer needs.

• Qwen3 comprises text-only dense transformer models in sizes of 0.6B, 1.7B, 4B, 8B, 14B and 32B, as well as MoEs in sizes of 30B-A3B and the flagship Qwen3-235B-A22B. All Qwen3 models are offered in base, thinking and instruct variants.
  
  
• Qwen3-Next is an experimental text-only MoE with 80B parameters (3B active) that replaces standard attention with Gated Delta Networks (inspired by Mamba-2) and Gated Attention.
  
  
• Qwen3-Omni is a natively multimodal model built on Qwen3-30B-A3B, supporting text, image, audio or video inputs and text or speech outputs.
  
  
• Qwen3-Coder-Next is a version of Qwen3-Next fine-tuned for code generation.
  
  
• Qwen3.5, released in February 2026, is a family of multimodal models utilizing the architecture first introduced in Qwen3-Next. It comprises both base and hybrid reasoning models in sizes of 0.8B, 2B, 4B, 9B and 27B, as well as MoE models in sizes of 35B-A3B, 122B-A10B and the flagship 397B-A17B. Qwen3.5-397B-A17B aims to compete with leading Gemini, GPT and Claude models for frontier performance.