Models

The Big Three

Frontier

• ChatGPT — OpenAI’s flagship. Broadest feature set in one place: voice, vision, images, code, apps. GPT-6 (7 October 2026) runs chat on GPT-6 Sol for Plus, Pro, Business and Enterprise and on GPT-6 Luna for Free and Go. It adds Intelligent UI: answers can include charts, buttons, forms and small working tools. ChatGPT Edu is not named in the rollout. Work and Codex still run GPT-6.1 Sol (29 September 2026), which OpenAI says nearly matches GPT-6 Astra at a fifth of the price. GPT-6.1 Terra (26 October 2026) is a faster, cheaper API model at $1/$5 per million tokens. On 11 December 2026 OpenAI retires GPT-5, GPT-5 mini, nano and pro, o3 and o3-pro from the API. Speed and quality claims are OpenAI’s own.
• Claude — Anthropic’s model. Strong on long documents and writing; Cowork and Code for agentic work.
• Gemini — Google has focused on deep integration with Google products, and is less generally capable except for best-in-class image generation. Gemini 4 Argon (Oct 13, 2026) for Google AI Pro and Ultra subscribers and through the API; Google says more Gemini 4 models will follow. Gemini 4 Carbon (Oct 28, 2026) is limited to selected cyber defenders for now. Argon’s introductory API price of $2/$10 per million tokens rises to $4/$20 on Dec 1.

Checked 2026-07-24

Other Proprietary Models

Frontier

• Grok — Designed with extreme personalities and a political bias. Real-time access to X.
• Copilot — Microsoft’s ChatGPT-based assistant. If your institution provides it in MS Office you may be FERPA and HIPAA secure.
• Ernie — Baidu’s multimodal frontier model. Ernie X1 is the reasoning version.
• Sonus — Family of proprietary models: Pro, Air, Mini, and Pro with Reasoning.
• Amazon Nova — Amazon’s model family, available through Bedrock.
• MAI-Thinking-1 — Microsoft’s first in-house model. Medium-size, built for math, coding and enterprise use.
• Muse Wonder — Meta’s frontier models (Oct 2026). Muse Wonder through the Meta Model API and Muse Code; Muse Wonder Full only for Muse Code subscribers. Weights not released. Benchmarks are the company’s own.
• Meta AI — Meta’s assistant, running Meta’s proprietary Muse Spark models since April 2026. Requires a login.
• Mercury 3 — Inception Labs’ diffusion model (October 2026), which writes a whole answer at once instead of word by word, at over 1,000 tokens a second. Inception says it matches or beats GPT-6 Luna and Claude Haiku 5.5 while running four to eight times faster. Speed and benchmark claims are the company’s own.

Education

• EduGPT — Built for security, privacy and education use. One of a series of EU models from CivAI.

Privacy

• Euria — Swiss Infomaniak model. Privacy-focused, renewable powered, recovered energy heats homes.
• Thaura — Free model from two Syrian engineers. Dedicated to global justice; no training on your data, no military contracts.

Conversation

• Pi — Focused on emotional intelligence, dialogue and role-playing. Talks but cannot hear.

Search

• You.com — Started as a privacy-focused search competitor; now focused on specialized API.

Checked 2026-07-24

Open Weight Models

All of these are open-weight models: you can download and run the weights. Few are fully open source; most do not release their training data or code.

Frontier

• Inkling — From Mira Murati’s Thinking Machines. Mixture-of-Experts, 1M token context.
• Kimi — Huge free multimodal Chinese model with a very large context window. Strong on math and coding. Full K3 weights released July 2026 — the largest open-weight model ever. Kimi K3.1 Pro (October 2026) is in the app now; Moonshot says its weights follow by November 4.
• Z.ai (GLM) — Beijing’s GLM series rivals Western proprietary models at lower cost. Usable without logging in. GLM-5.3 (August 2026) runs on the same base model as 5.2 — every gain came from more post-training, which is worth knowing if you think bigger always means better. Weights promised after safety testing.
• DeepSeek — Strong text and math; built for a fraction of the cost through Multi-head Latent Attention. Free to download.
• Qwen — Alibaba’s open-weight range with a slider for reasoning level. Guest mode available. In August 2026 Alibaba opened its top-tier model as well: Qwen3.8-Max, 2.4 trillion parameters with 95 billion active and a 1M-token context. The benchmark claims are Alibaba’s own.
• Muse Spark Open — Meta’s open-weight model (Oct 2026): 816B parameters, 42B active, 1M-token context; text, image, audio and video in. Gated download under Meta’s own license. Meta’s reference setup is two 512 GB M5 Ultra Mac Studios; 3-bit community versions run on one 512 GB Mac. Benchmarks are the company’s own.
• Beam — Reflection AI’s open-weight model for coding and agent work (Oct 2026): 501B parameters, 23B active, Apache 2.0. Free to start on Reflection’s API and on partner clouds. Benchmarks are the company’s own.
• Mistral Vibe — Mistral’s chat and agent app (formerly Le Chat). Fast, more multilingual than the big three, with press-wire news access. Most of its models are Apache 2.0. Large 4 (open weights, October 2026) uses a modified MIT license; companies that sell it as a hosted service need a separate license.
• MiniMax — Chinese company with open-weight reasoning models (M1), video and agent tools.
• Deep Cogito — Open-weight reasoning models. Current release is Cogito v2.1, a 671B model finetuned from DeepSeek.
• HuggingChat — Chatbot running on Llama. A good place to see what open-weight models can do. No login required.
• Nemotron 3.5 Lightning — NVIDIA’s 30B open-weight model, built to be the fast cheap worker an agent calls ten thousand times, not the brain that plans the job. Weights, training data and recipes all released.
• Muse Glimmer — Meta’s 30B open-weight model (Aug 2026) for running agents and coding on your own computer with one consumer GPU. Apache 2.0. Benchmarks are the company’s own.
• Nemotron 3.5 Super VL and Bolt — NVIDIA’s October 2026 additions to the family: Super VL (120B) reads documents, charts and screenshots for agent work; Bolt (74B, 9B active) fits in 24 GB of memory to run agents locally. Open weights, data and recipes. Bolt E4B (October 23, 2026) is built on Google’s Gemma 4 and runs in about 4 GB of memory.
• K2 Horizon — Six models from 0.9B to 375B, released 3 September 2026 by MBZUAI’s Institute of Foundation Models in Abu Dhabi under Apache 2.0, with weights, code and training data. Benchmark claims are the institute’s own.
• Kolibri — Aleph Alpha’s open-weight German and English model, released 3 October 2026 under Apache 2.0. A 78B mixture-of-experts model with 3.46B parameters active per token, a 262K-token native context that extends to 1M tokens, and tool calling. Built for government and regulated industries. Benchmark claims are the company’s own.

Reasoning

• MiMo — Xiaomi’s open-weight reasoning model.

Checked 2026-07-24

Regional & Culturally-Specific

North America

• Latimer — Named for Lewis Latimer. Trained further on licensed books, oral histories and sources from Black and Brown communities.

Latin America

• LatamGPT — Open-weight model from Chile’s CENIA trained on characteristic data from Latin America.
• Maritaca AI — Brazilian Portuguese chatbot and the Sabia family of fine-tuned models.
• Rio 3.5 Open 397B — Open-weight Portuguese and English model from IplanRIO, the City of Rio de Janeiro’s IT company, under an MIT license. Built by merging two models and distilling from a stronger one; the model card says the weights first uploaded were the wrong version and a correction is pending.

Middle East

• Fanar — Culturally and regionally aware Arabic LLM from Qatar’s MCIT and QCRI, fluent in Arabic dialects.
• Falcon — UAE open-weight model using state space architecture instead of transformers. Handles many dialects.
• Jais — Trained on the largest Arabic dataset used for an open-source foundational model. Inception rebranded to Inception42; Jais 2 released with Cerebras and MBZUAI.
• Qalam — Arabic writing and translation tool with plagiarism detection and sentiment analysis.
• Humain — Saudi model in Arabic and English, working with Qualcomm. Also makes an AI laptop.
• PersianLLaMA — 13B model fine-tuned on Persian Wikipedia.
• Karnak Chat — Egyptian government chat assistant built on Karnak, the national Arabic model from the Ministry of Communications’ Applied Innovation Center. Handles Modern Standard Arabic and Egyptian colloquial. Beta; sign-in required.
• K2 Horizon — Six models from 0.9B to 375B, released 3 September 2026 by MBZUAI’s Institute of Foundation Models in Abu Dhabi under Apache 2.0, with weights, code and training data. Benchmark claims are the institute’s own.

Europe and Russia

• Soofi S — German open-weight model funded by the Federal Ministry for Economic Affairs. Industrial and enterprise focus.
• AMALIA — Fully open-source 9B model natively trained in European Portuguese. Launched July 2026.
• Apertus — Swiss National Supercomputing Centre. 1000 languages, 40% non-English data, fully open development.
• OpenLLM-Ro — Set of LLaMA fine-tuned Romanian models.
• Bielik — Open-source family with deep integration of Polish language and cultural context.
• GigaChat — Open source Russian language model. Useful as a cultural model even where others score higher on MERA.
• UCCIX — Irish language LLM.
• Latxa — Family of LLaMA-based Basque models.

Africa

• Vulavula — Speaks, transcribes and translates low-resource African languages including isiZulu, Sesotho and African French.
• EthioLLM — Trained on Amharic, Ge’ez, Afan Oromo, Somali, Tigrinya and English.

South Asia

• Airavata — 7B model optimized for Hindi, from AI4Bharat.
• Nanda — Hindi-English open-weight model from G42 and MBZUAI. NANDA 87B, built on Llama 3.1 and trained on 65 billion Hindi tokens, released December 2025. No public model page could be found.
• Sarvam — India’s flagship AI lab: models built for Indian languages with India-hosted inference. A trillion-parameter model is announced but not yet released.
• BharatGen — India’s government-funded multimodal LLM initiative, fluent in 22+ Indian languages. Anchored by IIT Bombay.
• Bodhan AI — Open-weight models for speech recognition, text-to-speech, translation and OCR in Indian languages, from Bodhan AI, an IIT Madras-backed non-profit (September 2026).

Southeast Asia

• SEA-Lion — Open source LLM for Southeast Asian languages: Indonesian, Thai, Vietnamese, Filipino, Burmese, Khmer, Malay, Tamil, Lao.
• Komodo — First LLM built on Indonesian languages including Banjarese, Sundanese, Balinese and Toba Batak.
• Typhoon — Set of open source models optimized for Thai.

East Asia

• Doubao — ByteDance model with voice mode; one of the most popular AIs in China.
• Hy — Tencent’s Hy3 (July 2026, formerly Hunyuan) is an open-weight MoE: 295B total parameters, only 21B active, 256K context, Apache 2.0 licensed.
• TAIDE — 7B model grounded in Taiwanese culture, language, values and customs.
• Fugaku-LLM — Japanese-language model built on the RIKEN supercomputer Fugaku. Unrelated to Sakana’s Fugu.
• HyperCLOVA — Naver’s model, designed to understand Korean in its societal context.

Central Asia

• Sherkala — Pre-trained on Kazakh and English with some Russian and Turkish.

Checked 2026-07-24

Science & Math

Research

• Edison Platform — Single environment for scientific R&D. Commercial spinout of FutureHouse with multiple agents.
• Kosmos — Lab-in-the-loop AI scientist. One run reads ~1,500 papers and executes ~42,000 lines of code, fully auditable.
• Lila — Scientific AI connected to autonomous laboratories. Trained on 10 trillion tokens of scientific reasoning data.
• GPT-Rosalind — OpenAI’s model for life sciences research. Access requires approval; since Oct 2026, labs at US universities can apply for up to 30 free ChatGPT Business seats for up to 12 months. GPT-Rosalind-6 (30 Oct 2026) uses the same weights as GPT-6 Astra, with trusted access only.

Earth Science

• OlmoEarth — Foundation models for Earth observation and remote sensing: mangrove change, forest loss, crop-type maps.

Math

• WolframAlpha — Combines Wolfram Alpha’s computational power with ChatGPT.
• AlphaGeometry 2 — Competes at gold-medal IMO level. Not generally available.
• Numina-Lean-Agent — Uses standard models to do formal math reasoning.
• Mathstral — Built on Mistral 7B for mathematical reasoning. Open weights under Apache 2.0. Deprecated by Mistral and no longer updated.
• Aristotle — Harmonic’s theorem prover. Takes a problem in English, formalises it, and returns a machine-checked proof in Lean; it can also work inside an existing Lean project. Its IMO gold-medal result is reported in the company’s own preprint. Login required.

Physics

• Walrus — Trained on real scientific datasets rather than text. Domain is fluids and fluid-like systems.
• MatterChat — Berkeley Lab bridge model connecting an LLM with physics-based interatomic potentials.

Astronomy

• AION-1 — Trained on data from astronomical surveys. Polymathic AI collaboration with Cambridge.

Chemistry

• Ether0 — 24B reasoning model built on Mistral-Small-24B, post-trained for chemistry.

Materials

• SCIGEN — MIT model steering generation toward materials with quantum structures like Kagome lattices.

Checked 2026-07-24

World Models

A world model generates a space you move through rather than a clip you watch: you steer a camera and the model renders what comes next, keeping objects where you left them. That persistence is the difference from a video generator. Access is the current constraint – only one of these three is open to subscribers today.

Frontier

• Atlas — World Labs’ omni world model: generates, reconstructs and simulates 3D scenes from text, images, video or 3D data. Early access by request; World Labs says it will power future versions of Marble.
• Solaris — Runway’s interface world model: it draws an interactive app frame by frame as you use it, with no code underneath. Early access by request.
• Genie 3 — Google DeepMind’s world model, reached through Project Genie: a text or image prompt becomes an environment you walk through in real time. Requires a Google AI Ultra subscription and a US account. Genie 4 (Oct 27, 2026) adds sound and text steering, in a research preview for trusted testers.
• Marble — World Labs’ world model. Generates an explorable 3D scene from text, an image, a video or a rough 3D layout. Its Chisel tool blocks out scene structure in simple shapes first, then applies style separately.

Checked 2026-09-03

Historical LLMs

Historical

• Mr Chatterbox (Victorian) — Trained from scratch on 28,000 Victorian-era British texts, 1837-1899, from the British Library.
• TimeCapsuleLLM — Student-built model trained on 1800-1875 London texts to speak Victorian.
• Talkie — 13B model trained only on pre-1931 public domain text.
• MonadGPT — Trained on 11,000 texts from 1400-1700 using 17th-century knowledge frameworks.
• Latin BERT — Contextual language model for Latin.
• XunziALLM — Uses the formal rules of classical Chinese poetry.
• CuneiScribe — Reads Akkadian clay tablets and writes new ones. Type English, get cuneiform rendered as a tablet image, and it tells you when it is guessing. Published as TabletCraft (arXiv 2608.02609). Its own Gilgamesh demo mistranslates the famous opening line.

Checked 2026-07-24

Small Models & Edge AI

Small

• Phi-4 — Microsoft’s small open-weight family, with a reasoning variant. Supersedes Phi-3.5.
• OpenELM — Apple’s small model family, 270M-3B parameters.
• Gemma — Google’s open source smaller model, several sizes. Not the same as Gemini.

Checked 2026-07-24

Classifiers & Decision Models

These don’t write. They take text in and return a decision (classify, route, score, pick) with a confidence score, fast and cheap enough to sit inside a workflow. The waiter, not the chef.

Classifiers

• Jev — TypeSafe AI’s first “System One” model (Sept 2026). No text output: typed decisions with probabilities. Claims 193x faster and 445x cheaper than frontier models on its own tests; early independent tests are small but favorable, and one found its raw confidence scores overconfident.
• OpenAI Decisions API — OpenAI’s decision endpoint (DevDay, 29 September 2026). You set the questions and the allowed answers, send text or images, and OpenAI’s Luna model returns the answers, to classify content, route requests or choose an agent’s next step. Limited preview; broad release promised in the coming days.
• Microsoft-Decision-1 — Microsoft’s decision model (Oct 2026), built on Qwen3.5-9B. You give it the allowed answers and it returns a probability for each, for routing, classifying or grading against a rubric. No text output. In Microsoft Foundry and OpenRouter at $0.042 per million input tokens.
• Mammoth-1 and Elephant-1 — Base Labs’ decision models (October 2026), 27B and 9B, that return a probability for each allowed answer instead of text. Free to download under an MIT license; scores are the company’s own.
• Clairic — Cloudflare’s extraction model (preview, October 2026) that pulls values out of text, computes them, estimates what the text only hints at and forecasts the next values in a series. It writes no tokens; each answer comes with a confidence score. Open weights promised later.

Checked 2026-09-22

How Models Are Scored

Benchmarks

• EpochAI — Independent tracking of models and capabilities, with a benchmarks dashboard and a dataset of notable models.
• Arena — Head-to-head leaderboards, including text-to-image. Renamed from LMArena, formerly Chatbot Arena.
• MERA — Leaderboard for Russian-language tasks. A good example of why English-only benchmarks mislead.
• TasteVal — Measures how well a model designs experiments and interprets results on eight AI research tasks, scored against human experts by compute efficiency (Oct 2026). Results are the authors’ own; the tasks are not public.

Tracking

• Ithaka S+R Product Tracker — A longer list of AI products marketed to faculty and students.
• Emerging Trends in AI Agent Evaluation — Princeton-led survey (October 2026) of 33 trends and open questions in how AI agents are tested, from methods to who does the testing.

Checked 2026-07-24