Language Models AI tools
Large Language Models (LLMs) are advanced AI systems that understand and generate human-like text. These models power a wide range of applications from conversational AI to content generation, code writing, and knowledge retrieval. They serve as the foundation for many modern AI applications.
61 verified AI-first sites in Language Models.
AI lab behind ChatGPT, Codex, and the GPT model family. GPT-6 Astra is its flagship for complex reasoning, coding, and computer use, with GPT-6.1 Sol (September 2026) for near-Astra work at a fifth of the price, GPT-6.1 Terra (October 2026) for fast visual reasoning and computer use, and GPT-6 Luna for low-cost, high-volume tasks. The API also covers image, speech, and realtime models, plus open-weight gpt-oss releases.
Pricing: ChatGPT: Free and Go tiers; Plus $20/month; Pro, Business, Enterprise, and Edu plans. API: pay per token.
OpenAI's AI assistant for chat, search, writing, analysis, and agentic tasks. Voice, vision, file uploads, memory, and connectors to work apps; paid plans add ChatGPT Work and Codex with GPT-6 models. One of the most widely used AI products, with hundreds of millions of weekly users.
Pricing: Free and Go tiers. Plus: $20/month. Pro, Business, Enterprise, and Edu plans.
AI safety company and creator of the Claude model family. The current lineup is Claude Fable 5.1 for demanding long-horizon agentic work, Opus 5.5 for agentic coding and knowledge work, Sonnet 5.5, and Haiku 5.5, each with a 1M-token context window. Offers the Claude API on its own platform and through AWS, Google Cloud, and Microsoft Foundry, plus enterprise deployments and research on responsible AI.
Pricing: Free tier. Claude Pro: $20/month. Enterprise: Custom solutions.
Anthropic's Claude assistant for chat, coding, analysis, and agentic tasks. Runs the Claude 5.5 family of Opus, Sonnet, and Haiku models, with Claude Fable 5.1 for the most demanding multi-step work. Also available through the Claude Platform API, with monthly API credits on Max and Team plans for building your own apps and agents.
Pricing: Free tier. Pro: $20/month. Max and Enterprise: See site.
Paris-based AI lab shipping open-weight and commercial models. Mistral Large 4 (October 2026) is a 1T-parameter multimodal mixture-of-experts model with 49B active parameters and a 1M-token context, released as open weights under a modified MIT license that requires a separate license from model-as-a-service businesses. Also offers Medium, Small, and the Devstral and Codestral coding lines, the Le Chat assistant, the La Plateforme API, and private or on-prem deployment for European enterprises and governments.
Pricing: Free API access available. Pay-as-you-go pricing for higher usage. Enterprise licensing available.
Meta's open-weight Llama language model family (Llama 3/4 and prior). Downloadable weights for self-hosting and fine-tuning via Hugging Face and community tooling. Existing Llama releases remain open, while Meta's newer open-weight releases come from Meta Superintelligence Labs as Muse Spark Open and Muse Glimmer.
Pricing: Open weights free under Meta Llama license; commercial terms vary by release. Hosted inference billed by third-party providers.
Enterprise AI company focused on private, secure deployment. Command generative models, Embed and Rerank for retrieval, plus Transcribe, Translate, and Parse models; North workspace and Compass search build on them. Merging with Germany's Aleph Alpha (announced September 2026, pending approval) to form a transatlantic sovereign AI company.
Pricing: Free tier for development. Enterprise: Usage-based pricing. Custom solutions available.
Google DeepMind - Alphabet's frontier AI lab behind Gemini, Gemma, Imagen, Veo, and scientific systems such as AlphaFold. Merged Google Brain and DeepMind; research spans multimodal models, agents, robotics, and safety. Public research, open models, and products that power Google Search, Gemini apps, and Cloud.
Pricing: Research and many demos free. Gemini and Cloud AI: see Google pricing.
German sovereign AI company building the PhariaAI stack and models for governments and regulated enterprises, with on-prem deployment and EU AI Act compliance built in. Kolibri (October 2026) is a 78B-parameter German and English mixture-of-experts model with open Apache 2.0 weights and a 1M-token context. Signed a definitive merger agreement with Cohere in September 2026; the combined company will operate as Cohere once approved.
Pricing: Custom pricing based on usage. Enterprise solutions available.
Human-centric language models and the Pi assistant. After the 2024 Microsoft talent deal, Inflection rebuilt around relational AI-updated Pi with voice, memory, and agent tools, plus Inflection AI Labs experiments such as Pi Journeys. Also licenses models for private enterprise deployments.
Pricing: Free Pi tier; enterprise licensing-see Inflection.
Chinese lab shipping competitive open-weight LLMs under the MIT license. DeepSeek-V4.1-Flash (September 2026, updated October 17) is the top released open-weight model on the Artificial Analysis Intelligence Index, alongside the larger V4 Pro line. Chat app, API, and self-hosted checkpoints, often tuned for domestic Ascend as well as NVIDIA stacks.
Pricing: Free tier and API per-token pricing; see DeepSeek.
Alibaba's Qwen model family and chat app. Qwen3.8-Max (August 2026) is a 2.4T-parameter multimodal flagship with a 1M-token context and, for the first time at Max scale, open weights. Open Qwen3.5 and 3.6 sizes run from 0.8B to 397B; API via Alibaba Cloud Model Studio.
Pricing: Custom pricing based on usage. Enterprise solutions available.
Multimodal language model platform that processes streams of video, audio, and text with models trained from scratch. Offers vision, speech, and chat capabilities with on-device deployment options. Rho-1 (October 2026, research preview) is a 19B omni-reasoning model that understands and generates text, images, and video in one network.
Pricing: Contact for pricing
NVIDIA's Nemotron family of open models, released with open weights, training data, and recipes for building reasoning agents. The Nemotron 3.5 line (October 2026) includes Super VL, a 120B-parameter model with 12B active, and Bolt, a 74B model with 9B active, alongside small variants. Weights are published on Hugging Face under NVIDIA's OpenMDW license.
Pricing: Open weights; free to download and modify.
Google's multimodal language model family (Gemini). Native text, image, audio, and video understanding with long context, API access via Google AI Studio and Vertex, and consumer apps on gemini.google.com. Gemini 4 Argon (October 2026) is the current frontier model in the API, the Gemini app for AI Pro and Ultra subscribers, and AI Mode in Search, while the stronger Gemini 4 Carbon is limited to selected cyber defenders through the Fairwind Program.
Pricing: Free tier available. Premium subscriptions available. API pricing varies by usage.
Frontier lab building open-weight foundation models for software engineering (Laguna family). Laguna S 2.1 (July 2026) is a 118B mixture-of-experts model with 8B active parameters and a 1M-token context under OpenMDW-1.1, alongside Laguna XS 2.1 (33B, 3B active) and Laguna M.1 (225B, 23B active). Available through the Poolside API, OpenRouter, and pool, its terminal coding agent.
Pricing: Enterprise pricing. Available via Amazon Bedrock and EC2. Contact for custom deployment.
Company building ultra-efficient foundation models (LFMs) for on-device and edge deployment. LFM3 (October 2026) puts text, vision, audio, omni, and decision models on one shared backbone in the roughly 2B-parameter range, with open weights on Hugging Face, while the LFM2.5 family spans 350M to 24B parameters. Runs on GPUs, CPUs, and NPUs, with the LEAP platform for customization and deployment and the Apollo app for private on-device chat; used by Shopify, Qualcomm, AMD, and Capgemini.
Pricing: Free for companies under $10M revenue. Startup program and enterprise: contact for pricing.
Moonshot AI's Kimi assistant and model platform. Kimi K3.1 Pro (October 2026) has about 3T parameters and a 1M-token context for long-horizon coding, 3D creation, financial research, and agent tasks. Includes long-context chat, deep research, the Kimi Code agent, open-weight K2 and K3 releases, and an API at platform.moonshot.ai.
Pricing: Free tier. Starter: $9/mo (10M tokens). Ultra: $49/mo (70M tokens). API: pay-as-you-go. See platform.moonshot.ai for details.
Chinese AI company building language, speech, video (Hailuo), and music models. Its M3 flagship (June 2026) targets coding and agentic work at low API prices, and M3.1 Flash (October 2026) ships open weights under the MiniMax community license. Also offers the MiniMax Code assistant, with API and consumer products serving users in 200+ countries.
Pricing: API usage-based pricing. Enterprise solutions available.
Chinese foundation model provider behind the GLM model series, offering large language model APIs and enterprise AI capabilities for chat, reasoning, and multilingual use cases. GLM-5.3 is an open-weight mixture-of-experts model with about 753B parameters, 40B active, and a 1M-token context, ranking second among released open-weight models on the Artificial Analysis Intelligence Index. The Z.ai chat and agent app runs GLM-5.3-Flash.
Pricing: API and enterprise pricing; see site for current options.
Baidu's large language model family for conversational AI, content generation, and enterprise AI applications. ERNIE 5.1 (May 2026) is its current flagship text model, hosted-only with a 128K-token context through ERNIE Bot, AI Studio, and the Qianfan API. Baidu's open-weight line remains ERNIE 4.5 under Apache 2.0.
Pricing: Platform and enterprise pricing; see Baidu/ERNIE pages for current details.
AI company building foundation models and document intelligence for enterprise. Its Solar family includes Solar Open 2 (July 2026), a 250B-parameter mixture-of-experts model with 15B active parameters built for agents and long documents, developed under Korea's sovereign AI foundation model project. Also offers APIs and solutions for text generation, classification, and document understanding.
Pricing: API and enterprise pricing; see upstage.ai for current plans and trials.
AI lab building diffusion-based reasoning language models for production latency. Mercury family (including Mercury 2) refines full responses in parallel instead of token-by-token decoding, targeting reasoning-grade quality at 1000+ tokens/sec on NVIDIA Blackwell-class hardware. OpenAI-compatible API with tunable reasoning, 128K context, native tool use, and schema-aligned JSON. Distinct from the Allen Institute's OLMo open models and from hyperscaler frontier APIs.
Pricing: API from about $0.25/1M input and $0.75/1M output tokens; see inceptionlabs.ai.
Open language model program from the Allen Institute for AI (AI2). Releases fully open OLMo 2 and related foundation weights, training data, evaluation harnesses, and recipes-not just model checkpoints. Built for reproducible research and commercial use under open licenses; complements AI2's separate Asta science-agent products. Models span efficient to large scales with transparent pretraining and ongoing community releases on Hugging Face and GitHub.
Pricing: Open-source models and artifacts free; hosted inference via third-party providers.
Open-model research lab behind the Hermes LLM family for agentic, reasoning, and roleplay workloads. Hermes 3 and 4 open-weight releases with Nous Portal API and Hermes Agent product. Also builds Nomos and Psyche models; strong community adoption for fine-tunes and agent frameworks.
Pricing: Nous Portal subscription and API; open weights free to download.
Frontier AI lab founded by Mira Murati. Inkling (July 2026) is its open-weights multimodal mixture-of-experts model with 975B total and 41B active parameters, a 1M-token context, text, image, and audio input, and an Apache 2.0 license, with the smaller Inkling-Small alongside it. Tinker, its training platform, fine-tunes open models including Inkling, Qwen, and Kimi through a managed API.
Pricing: Open weights free under Apache 2.0; Tinker fine-tuning and API billed by usage.
Chinese foundation-model lab (阶跃星辰) behind the Step family of language, vision, audio, and video models. Step 5 Preview is an open-weight 600B mixture-of-experts model with 27B active parameters and a 1M-token context, alongside Step-3.7 Flash for production agent workflows with tool calling and MCP compatibility. OpenAI-compatible API at platform.stepfun.com; backed by Tencent and Shanghai government.
Pricing: API pay-as-you-go; Step Plan subscriptions; see platform.stepfun.com.
Meituan's LongCat chat and API for their open LLM family, including LongCat-2.5 (November 2026) with 1.6T total parameters and 48B active. The multimodal model uses a unified token architecture across text and vision for agent execution and extended programming workflows. Consumer chat at longcat.chat plus OpenAI- and Anthropic-compatible endpoints for agentic coding, reasoning, and long-context work.
Pricing: Free chat tier; API via LongCat platform keys-see site for limits and pricing.
Meta Superintelligence Labs' Muse model family, including the proprietary Muse Spark 1.3 and the Muse Wonder and Muse Wonder Full models released in October 2026. Muse Spark Open (October 15, 2026) is the open-weight member, with 816B total and 42B active parameters and a 1M-token context. Multimodal reasoning for agentic coding, long-running agents, tool use, and computer use, available in Meta AI, in Muse Code, and via the Meta Model API.
Pricing: In Meta AI apps; Model API pay-as-you-go-see Meta developer docs.
Microsoft's open Phi small language model family for efficient on-device and cloud reasoning. Phi-4-reasoning-vision-15B (March 2026) is the latest broadly released model, a 15B open-weight multimodal reasoning model under the MIT license for math, science, document understanding, and screen grounding. Available through Microsoft Foundry, Hugging Face, and GitHub.
Pricing: Open weights and Azure-hosted options-see Microsoft Phi product pages.
IBM's Granite foundation model family for enterprise language, code, and multimodal workloads. Granite 4.1 (April 2026) provides dense 3B, 8B, and 30B language models trained on about 15T tokens, plus vision, speech, embedding, and Guardian models, all under Apache 2.0. Granite Switch 4.1 (October 2026) composes a Granite base model with Decision and Guardian adapters and speech-to-text behind one endpoint.
Pricing: Open models and watsonx commercial options-see IBM Granite.
ByteDance Seed foundation-model lab behind Seed LLMs and related multimodal systems powering Doubao and ByteDance AI products. Seed 2.0 (February 2026) offers Pro, Lite, and Mini general agent models plus a Code model, with APIs on Volcano Engine and use in the Doubao app and TRAE. The same lab builds the Seedance video models, with Seedance 2.5 launched in July 2026.
Pricing: Product and API access via ByteDance / Doubao channels-see Seed site.
Tencent's Hy model family, renamed from Hunyuan in December 2025, covering large language, vision-language, and visual generation models. Hy4 preview was open-sourced with 770B parameters, 49B active, and a 1M-token context, and Hy4 (November 4, 2026) adds vision with gains in long-horizon tasks and deep research. Available through Tencent Cloud, OpenRouter, and Tencent products such as Yuanbao, CodeBuddy, and WorkBuddy.
Pricing: Tencent Cloud API and product pricing-see Hunyuan.
Meta Superintelligence Labs' open-weight Muse Glimmer 30B - Apache 2.0 multimodal model built for always-on local agents. Runs on a single consumer GPU or Mac; strong tool use, long-running tasks, and coding versus peers in its size class. Distinct from closed Muse Spark; weights on Hugging Face with Ollama, LM Studio, and cloud hosts following.
Pricing: Open weights free under Apache 2.0; self-host hardware costs; hosted inference billed by third parties.
Google DeepMind's open-weight Gemma model family (Gemma 4). Dense and MoE sizes from edge (E2B/E4B) to 31B for reasoning, coding, and multimodal work. Apache 2.0 weights; the open counterpart to Gemini for self-host and fine-tunes.
Pricing: Open weights free (Apache 2.0); hosted inference billed by Google and third parties.
Amazon's Nova foundation model family on AWS for text, multimodal, and agent workloads. Nova 2 Lite is a reasoning model with a 1M-token context, configurable thinking effort, a built-in code interpreter and web grounding, and remote MCP tool support. The family also includes Nova 2.5 Sonic for speech-to-speech voice agents, Nova Multimodal Embeddings, and Nova Forge for building custom frontier models.
Pricing: Pay-per-token on Amazon Bedrock; see AWS Nova pricing.
OpenAI's open-weight reasoning models (gpt-oss-120b and gpt-oss-20b) under Apache 2.0. MoE checkpoints for agents and private serving; native MXFP4 so 120b fits an 80GB GPU and 20b fits ~16GB. Open counterpart to GPT APIs.
Pricing: Open weights free (Apache 2.0); self-host hardware or third-party hosts.
Shanghai AI Laboratory's InternLM and Intern family of open language models. Intern-S2 (September 2026) is a 397B-scale scientific multimodal mixture-of-experts model with a 256K-token context, configurable thinking, tool calling, and Apache 2.0 weights. The lab also open-sources InternVL multimodal models and offers a hosted API with a free quota.
Pricing: Open weights free; hosted inference via third-party providers.
Elon Musk's frontier AI lab, now part of SpaceX and branded SpaceXSI (previously SpaceXAI) at x.ai. Builds the Grok family of language models; Grok 4.7 (September 2026) is its most capable reasoning model, trained on its Colossus cluster. Offers one API for text, code, voice, image, and video, plus Grok Build and Grok Bot for agentic work.
Pricing: Grok free and SuperGrok tiers; API pricing on xAI.
Consumer chat interface for the Grok models from xAI, now branded under SpaceXSI. Grok 4.7 handles coding and knowledge work, with real-time context from X, image and video generation through Grok Imagine, and voice chat. Grok Bot adds shared AI teammates for teams, and Grok Build is an open-source agentic coding tool.
Pricing: Free tier; SuperGrok and X Premium+ for higher limits.
Technology Innovation Institute's Falcon open model family from Abu Dhabi. The line includes Falcon-H1 hybrid models, Falcon 3, Falcon Mamba, Falcon Arabic, and the Falcon Perception model. Open-weight releases are widely used as a sovereign and research alternative to US closed labs.
Pricing: Open weights; see TII Falcon for licenses and downloads.
Apple's on-device foundation model framework for apps on Apple silicon. Local LLM and generation APIs powering Apple Intelligence features with privacy-preserving inference. Apple documents the Foundation Models framework at developer.apple.com for adding the on-device model to apps.
Pricing: Included with supported Apple platforms for developers; see Apple docs.
Inflection's personal AI companion chatbot. Conversational assistant focused on helpful, emotionally aware dialogue rather than enterprise coding agents. Inflection describes its mission as personal intelligence, empowering people with human-centered, emotionally intelligent AI.
Pricing: Free and paid tiers; see Pi.
Chinese foundation-model lab (百川智能) behind the Baichuan open and commercial LLM families. Models and APIs for chat, reasoning, and enterprise Chinese-language applications. Its stated mission is to give people easy access to world knowledge and professional services through breakthroughs in language AI.
Pricing: API and enterprise plans; see Baichuan.
SenseTime's SenseNova multimodal large-model platform. Native multimodal LLMs and API services for Chinese enterprise and consumer applications across text, vision, and agents. The Chinese-language platform advertises text-to-image, code generation, and conversational AI with enterprise-grade security and instant access.
Pricing: Platform and API plans; see SenseNova.
Open RNN-style language model family with Transformer-like quality and linear-time inference. A 100% attention-free Linux Foundation AI project combining parallelizable training with constant memory and unbounded context. RWKV-7 Goose reasoning models are the latest line, with GGUF and Ollama weights, an RWKV App for Android, iOS, and desktop, and inference tooling such as Albatross and RWKV-Runner.
Pricing: Open weights and tooling; see RWKV docs and Hugging Face.
Moonshot AI lab behind the Kimi model family. Kimi K3.1 Pro (October 2026) has about 3T parameters and a 1M-token context, following Kimi K3 in July 2026. Offers research releases, an API platform, and consumer Kimi chat products for long-context reasoning, coding agents, and deep research.
Pricing: API and product plans via Moonshot / Kimi; contact Moonshot.
US lab building open-weight foundation models. The Trinity family is released for anyone to inspect, fine-tune, and self-host, with the Arcee Platform for deploying and operating open models in production. Developing Genesis-Science-1 with the Department of Energy; valued at $1B+ after its 2026 Series B.
Pricing: Open weights free; Arcee Platform and enterprise plans via sales.
Beijing lab publishing Naive-N0.5-Flash, an open-weight 309B mixture-of-experts language model with 15.5B active parameters and a native 1M-token context for coding and AI research. It was built on Xiaomi's MiMo-V2.5 base with a new hybrid sliding-window and sparse attention design and about 3T tokens of continued training. Weights and inference code are released under the MIT license, along with the NaiveRT runtime that reaches up to 2,000 tokens per second.
Pricing: Open weights under MIT license; announced API pricing of $0.10 input and $0.40 output per million tokens.
MBZUAI research lab building fully open foundation models, with teams in Abu Dhabi, Silicon Valley, and Paris. Its K2 Horizon family (September 2026) spans six sizes from 0.9B to 375B-A23B parameters and ships under Apache 2.0 with intermediate checkpoints, training data or recipes, code, and logs for every stage. It continues the LLM360 open-model line and has day-zero support in vLLM, SGLang, and Ollama.
Pricing: Open weights under Apache 2.0; free to download from Hugging Face.
Xiaomi's MiMo model lab, which open-sources its own large language models. MiMo-V2.6 (September 2026) comes as Pro and Flash natively multimodal models, with the Pro scoring 46 on the Artificial Analysis Intelligence Index, plus a 9B Distill-Qwen variant, a technical report, and RL training resources. Weights are published on Hugging Face under the MIT license and served through the Xiaomi MiMo Open Platform API.
Pricing: Open weights under MIT license; pay-as-you-go API on the Xiaomi MiMo Open Platform.
Meta Superintelligence Labs' first frontier-class model series, released October 12, 2026 as muse-wonder-1.0 and Muse Wonder Full. Built for frontier coding, long-running agents, and computer use, it plans long tasks, delegates to sub-agents, and recovers from errors, with an Active Memory feature that carries context across conversations. Both models have a 1M-token context window; Muse Wonder is on the Meta Model API and in Muse Code, while Wonder Full is limited to Muse Code subscribers on the Power and Maximum Usage plans.
Pricing: Meta Model API $4 input / $0.20 cached / $20 output per million tokens; Wonder Full included with Muse Code Power ($50/month) and Maximum ($200/month) plans.
Open-weight 40B mixture-of-experts coding model from Recursive Superintelligence, with about 2.5B parameters active per token. Apache 2.0 weights on Hugging Face run locally through llama.cpp, Ollama, LM Studio, vLLM, and SGLang, plus Recursive's rec engine for Apple Silicon. Recursive says its Eureka Machine automated research system shaped architecture, training, quantization, and kernels, and reports leading scores among sub-50B MoE models on SWE-Bench Pro v2, Terminal-Bench 2.1, and SWE-Bench Verified.
Pricing: Open weights free (Apache 2.0); run locally. Hosted API terms separate.
Frontier lab building open-weight models and the software stack around them. Beam is its first model, a 501B mixture-of-experts model with 23B active parameters, a 1M-token context, and an Apache 2.0 license, built for coding, reasoning, and agentic workloads. Models are served through its API platform or deployed in a customer's own environment, from private cloud to air-gapped infrastructure.
Pricing: Open weights free (Apache 2.0); API platform and enterprise deployments via reflection.ai.
Indian sovereign AI lab that trains its own models from scratch. Sarvam 30B and Sarvam 105B are open-weight reasoning models under Apache 2.0, with the 105B a mixture-of-experts model with 10.3B active parameters and a 128K-token context, and both tuned for 22 Indian languages. Weights are on Hugging Face and AIKosh, and the models power its Indus assistant and are available through its API.
Pricing: Open weights free (Apache 2.0); API and Indus plans via sarvam.ai.
Ant Group's model program behind the Ling, Ring, and Ming series. Ling-3.1-flash is a 560B-parameter hybrid-linear mixture-of-experts reasoning model with 25B active parameters, a 1M-token context specification (256K on the official service), and MIT-licensed weights on Hugging Face. Ring is the deep-reasoning line, Ming handles omni text, image, audio, and video, and an open API serves the models.
Pricing: Open weights free; open API on a limited-time offer, see ant-ling.com.
Open superintelligence lab and model company based in San Francisco. ZAYA1-8B is a mixture-of-experts reasoning model with under one billion active parameters, trained entirely on AMD hardware and released under Apache 2.0, alongside ZAYA1-74B-Preview and ZAYA1-VL-8B. Models run as serverless endpoints on Zyphra Cloud and can be deployed in a customer's own environment. Its MAIA multiplayer agent is in beta, and the Customization Studio is still in development.
Pricing: Open weights free (Apache 2.0); Zyphra Cloud and enterprise plans via zyphra.com.
Microsoft AI's first reasoning model, trained from scratch by its MAI superintelligence team without distillation from third-party models. It is a mixture-of-experts model with about 1T total parameters, 35B active, and a 256K-token context window, aimed at software engineering and advanced math. Available through Microsoft Foundry, with OpenRouter, Fireworks, and Baseten distribution and tunable weights announced for developers.
Pricing: Available through Microsoft Foundry; see Microsoft AI for pricing.
Korean AI company building foundation models in-house. Motif 3 is a 314B-parameter mixture-of-experts language model with 13.2B active parameters, a 256K-token context, and MIT-licensed weights, built from scratch on about 12.5T tokens with a custom Grouped Differential Latent Attention design. The Motif chat app is free to try, and the company also offers smaller language, image, and video models plus on-premises and fine-tuning services.
Pricing: Free chat; open weights under MIT; on-premises and fine-tuning via contact.
LG AI Research's EXAONE family of foundation models. K-EXAONE 2.0 (July 2026) is a 750B-parameter mixture-of-experts model with 37B active parameters, a 262K-token context, 10 languages, and Apache 2.0 weights on Hugging Face, built under Korea's sovereign AI foundation model project. EXAONE 4.5 adds a native multimodal design with an integrated visual encoder for document and industrial vision tasks.
Pricing: K-EXAONE 2.0 weights free under Apache 2.0; see LG AI Research for other models.
SK Telecom's A.X family of Korean-specialized large language models. A.X K2 (July 2026) is a 688B-parameter mixture-of-experts model with 33B active parameters and a 256K-token context, trained from scratch with native FP8 and released under Apache 2.0. It tied for the top score on the MathArena AIME 2026 leaderboard at 97.1%, and A.X VE vision and A.X K2 audio derivatives are published alongside it.
Pricing: Open weights free (Apache 2.0); self-hosted.