Reference
Models
A directory of the models a reader of this guide is most likely to meet: the ones rented from a provider, and the ones whose weights can be downloaded and run. Search it, or filter by what a model does.
Every entry carries a link to where its facts came from, and says whether the model is current, has been passed by a successor, or has been switched off. This field moves fast: of the models here, several were retired during 2026, so check the date above and follow the source before relying on anything.
Search and filter
Showing all 92 models.
| Model | Provider | How you get it | Size | Released | Status | What it does | Details |
|---|---|---|---|---|---|---|---|
| ModelDeepSeek-V4.1-Flash-1017 | ProviderDeepSeek | How you get itOpen weights | Size552B total | ReleasedOctober 2026 | StatusCurrent | What it does | |
A post-training update to V4.1 Flash, served under the same API name. DeepSeek reports gains on all eleven benchmarks it lists, faster task completion and shorter outputs. Its Terminal-Bench score was measured with a different harness from the earlier one, so the two are not directly comparable. What it does
Released October 2026 Checked 8 November 2026
Where this came from for DeepSeek-V4.1-Flash-1017 (opens in a new tab) |
|||||||
| ModelGPT-6.1 Sol | ProviderOpenAI | How you get itHosted | SizeNot published | ReleasedOctober 2026 | StatusCurrent | What it does | |
OpenAI's workhorse tier, released on 2 October 2026 as a step up from GPT-6 Sol. OpenAI says it comes close to GPT-6 Astra on complex work, with better coding and computer use, at about a fifth of Astra's standard token price. Figures here come from third-party summaries, as OpenAI's model page could not be fetched. What it does
Released October 2026 Checked 8 November 2026
|
|||||||
| ModelGPT-6.1 Terra | ProviderOpenAI | How you get itHosted | SizeNot published | ReleasedOctober 2026 | StatusCurrent | What it does | |
OpenAI's mid tier, released on 26 October 2026 as the successor to GPT-5.6 Terra, aimed at computer use, visual reasoning and coding in ChatGPT Work and Codex rather than ordinary Chat conversations. OpenAI says it costs about half as much and runs faster than its predecessor, with fewer errors on its benchmarks. Figures here come from third-party summaries, as OpenAI's model page could not be fetched. What it does
Released October 2026 Checked 8 November 2026
|
|||||||
| ModelGrok Voice 3.0 | ProviderxAI | How you get itHosted | SizeNot published | ReleasedOctober 2026 | StatusCurrent | What it does | |
A speech to speech model for voice agents, which hears and answers in one pass instead of chaining transcription, a language model and speech synthesis. xAI claims under a second of latency. It is reached over a realtime WebSocket API shaped like OpenAI's, and the latest alias moved to it automatically. The figures so far are xAI's own and secondary coverage of them. What it does
Released October 2026 Checked 8 November 2026
Where this came from for Grok Voice 3.0 (opens in a new tab) |
|||||||
| ModelMuse Spark Open | ProviderMeta | How you get itOpen weights | Size816B total, 42B active | ReleasedOctober 2026 | StatusCurrent | What it does | |
Meta's first large open-weight model since Llama 4, after Muse Spark itself stayed behind an API. It is a mixture of experts model that reads text, images, video and audio and answers in text, aimed at coding agents and tool use. The licence carries commercial limits, and the repository was gated at release. The benchmark figures so far are Meta's own, and independent indices place it behind Kimi K3 and GLM 5.3. What it does
Released October 2026 Checked 8 November 2026
Where this came from for Muse Spark Open (opens in a new tab) |
|||||||
| ModelMuse Wonder | ProviderMeta | How you get itHosted | SizeNot published | ReleasedOctober 2026 | StatusCurrent | What it does | |
Meta's first closed frontier model, aimed at agent work. It is offered through Meta's Model API and its Muse Code product, and carries memory across conversations and a habit of asking before consequential actions. A larger sibling, Muse Wonder Full, is limited to Muse Code subscribers and is not in the API. The benchmark claims so far are Meta's own. What it does
Released October 2026 Checked 8 November 2026
|
|||||||
| ModelClaude Fable 5.1 | ProviderAnthropic | How you get itHosted | SizeNot published | ReleasedSeptember 2026 | StatusCurrent | What it does | |
The top of Anthropic's range, above Opus 5, for demanding reasoning and long-horizon agentic work. Its thinking is adaptive and always on, and cache reads cost a quarter of the usual rate. Anthropic's own advice is to start with Opus 5 and move here only when that model at higher effort still falls short. What it does
Released September 2026 Checked 8 November 2026
Where this came from for Claude Fable 5.1 (opens in a new tab) |
|||||||
| ModelClaude Opus 5.5 | ProviderAnthropic | How you get itHosted | SizeNot published | ReleasedSeptember 2026 | StatusCurrent | What it does | |
Anthropic's replacement for Opus 5, at the same context window but a fifth cheaper: Anthropic reports it costs about 40% less to run and answers more than 30% faster, while matching or beating Fable 5.1 on several agentic coding benchmarks. Thinking is always on and cannot be disabled, one of several breaking changes for code written against Opus 5. What it does
Released September 2026 Checked 8 November 2026
Where this came from for Claude Opus 5.5 (opens in a new tab) |
|||||||
| ModelClaude Sonnet 5.5 | ProviderAnthropic | How you get itHosted | SizeNot published | ReleasedSeptember 2026 | StatusCurrent | What it does | |
Anthropic's fast everyday tier, replacing Sonnet 5 at the same prices and generating output more than 30% faster by Anthropic's account. Thinking is adaptive, and a new setting turns off up-front thinking. It carries several breaking changes for code written against Sonnet 5: forced tool use returns an error, thinking blocks are tied to the model that produced them, and text between tool calls now comes back inside thinking blocks. What it does
Released September 2026 Checked 8 November 2026
Where this came from for Claude Sonnet 5.5 (opens in a new tab) |
|||||||
| ModelDeepSeek-V4.1-Flash | ProviderDeepSeek | How you get itOpen weights | Size552B total | ReleasedSeptember 2026 | StatusSuperseded | What it does | |
The small, fast model in DeepSeek's V4 family, rebuilt on a new causal encoder-decoder architecture and now reading images natively, which the V4 Flash it replaces could not. It supports tool calls, JSON output, and both an OpenAI- and an Anthropic-shaped API endpoint. The old model names still resolve, but requests to them are now served by this model and billed at its price. What it does
Released September 2026 Checked 8 November 2026
Where this came from for DeepSeek-V4.1-Flash (opens in a new tab) |
|||||||
| ModelGPT Image 2.5 | ProviderOpenAI | How you get itHosted | SizeNot published | ReleasedSeptember 2026 | StatusCurrent | What it does | |
OpenAI's image generation and editing model, replacing GPT Image 2 in ChatGPT, Codex and the API, with image generation latency cut by up to half. It comes in two API variants: Flare for fast, everyday generation, and Sunburst for slower, more precise work. ChatGPT also gained Sketch, which turns a drawing made in the app into a reference for the final image. What it does
Released September 2026 Checked 8 November 2026
|
|||||||
| ModelGPT-6 Astra | ProviderOpenAI | How you get itHosted | SizeNot published | ReleasedSeptember 2026 | StatusCurrent | What it does | |
OpenAI's flagship, at the top of its range for long work carried to completion: multi-step reasoning, software engineering, computer use and research. It takes text and images and returns text only, so it generates no media itself, though it can call image generation and other tools. What separates it from the GPT-5.6 tier is that it is built to finish a task over many steps rather than answer one turn well. What it does
Released September 2026 Checked 8 November 2026
|
|||||||
| ModelGPT-6 Sol | ProviderOpenAI | How you get itHosted | SizeNot published | ReleasedSeptember 2026 | StatusCurrent | What it does | |
OpenAI's workhorse tier below GPT-6 Astra, replacing GPT-5.6 Sol at half its price and behind most paid ChatGPT conversations. OpenAI reports about half as many factual mistakes as its predecessor on its internal accuracy evaluation, closing much of the gap with Astra. It arrived alongside a cheaper, faster sibling, GPT-6 Luna, sharing the same window; both keep a selectable reasoning effort, down to none for latency-sensitive calls. What it does
Released September 2026 Checked 8 November 2026
|
|||||||
| ModelGrok 4.7 | ProviderxAI | How you get itHosted | Size2.1T parameters | ReleasedSeptember 2026 | StatusCurrent | What it does | |
xAI's new flagship, built on a larger base model than Grok 4.6 and trained with a longer reinforcement learning run weighted toward problems that take many hours, aimed at coding, engineering and long office work. Its training mix was supplemented with SpaceX's own engineering records. It reaches Cursor, Grok Build, GitHub Copilot and the xAI API at once, without a waitlist, at an unchanged price. What it does
Released September 2026 Checked 8 November 2026
|
|||||||
| ModelMAI-Transcribe-2 | ProviderMicrosoft | How you get itHosted | SizeNot published | ReleasedSeptember 2026 | StatusCurrent | What it does | |
A speech to text model covering sixty languages with automatic language identification, so one model can serve a multilingual product. It adds speaker diarisation, word-level timestamps and keyword biasing for domain terms, and offers a verbatim or a cleaned-up style. What it does
Released September 2026 Checked 8 November 2026
Where this came from for MAI-Transcribe-2 (opens in a new tab) |
|||||||
| ModelMiMo-V2.6-Flash | ProviderXiaomi | How you get itOpen weights | Size309B total, 15B active | ReleasedSeptember 2026 | StatusCurrent | What it does | |
The smaller sibling of MiMo-V2.6-Pro, released with it on 21 September 2026 as open weights. It is a mixture of experts model that reads text, images, video and audio and answers in text, with only 15B parameters active per token, so it is far cheaper to serve than Pro, though still needing roughly 256GB of memory to run locally. What it does
Released September 2026 Checked 8 November 2026
Where this came from for MiMo-V2.6-Flash (opens in a new tab) |
|||||||
| ModelMiMo-V2.6-Pro | ProviderXiaomi | How you get itOpen weights | Size1.02T total, 42B active | ReleasedSeptember 2026 | StatusCurrent | What it does | |
Xiaomi's largest open model yet, and by its own benchmarks the strongest open-weights model at release, close to Claude Opus 5 and GPT-5.6 Sol on agent tasks. It reads text, images, video and audio, though it answers only in text, and comes from one mixed reinforcement learning run across coding, general agent, visual and cybersecurity tasks rather than a separate run per domain. A cheaper Flash sibling and an UltraSpeed variant, about twenty times faster, were released alongside it, and Xiaomi also opened over 7,000 of its RL training environments. What it does
Released September 2026 Checked 8 November 2026
|
|||||||
| ModelMAI-Code-1.1-Flash | ProviderMicrosoft | How you get itHosted | Size138B total, 5B active | ReleasedAugust 2026 | StatusCurrent | What it does | |
A small, fast coding model for everyday developer work rather than the hardest problems, and the model served inside GitHub Copilot and VS Code. It takes images as well as text, so a screenshot or a design can be part of the prompt. It was trained against the production Copilot harness itself. What it does
Released August 2026 Checked 8 November 2026
Where this came from for MAI-Code-1.1-Flash (opens in a new tab) |
|||||||
| ModelMAI-Cyber-1-Flash | ProviderMicrosoft | How you get itHosted | Size137B total, 5B active | ReleasedAugust 2026 | StatusCurrent | What it does | |
A security-specialised version of the Copilot coding model that finds, triages and patches vulnerabilities in large codebases. It is not sold on its own and runs only inside Microsoft's own scanning harness, handling the cheaper work and passing the hardest on. Because the capability is dual use, access is restricted to defenders who pass a review. What it does
Released August 2026 Checked 8 November 2026
Where this came from for MAI-Cyber-1-Flash (opens in a new tab) |
|||||||
| ModelMAI-Image-2.6 | ProviderMicrosoft | How you get itHosted | Size20B | ReleasedAugust 2026 | StatusCurrent | What it does | |
A diffusion model for generating images from text and for precise editing of images you already have, with multi-image reference editing. Microsoft reports stronger text rendering inside images than the previous version. A cheaper, faster sibling serves high-throughput work. What it does
Released August 2026 Checked 8 November 2026
|
|||||||
| ModelMAI-Thinking-1 | ProviderMicrosoft | How you get itHosted | Size35B active of about 1T total | ReleasedAugust 2026 | StatusCurrent | What it does | |
Microsoft AI's first reasoning model, text only, aimed at maths, coding and agentic enterprise work. Its distinguishing claim is that it was trained from scratch on licensed and traceable data without distilling from another lab's model, which Microsoft argues matters for provenance. It is in public preview on Microsoft Foundry. What it does
Released August 2026 Checked 8 November 2026
Where this came from for MAI-Thinking-1 (opens in a new tab) |
|||||||
| ModelQwen3.8-Max | ProviderAlibaba | How you get itOpen weights | Size2.4T total, 95B active | ReleasedAugust 2026 | StatusCurrent | What it does | |
The largest model Qwen has published, and the first of its Max tier released as weights rather than kept behind the API. Thinking mode is always on, so every reply begins with reasoning. The hosted version adds multimodal input the open weights do not have. What it does
Released August 2026 Checked 8 November 2026
|
|||||||
| ModelClaude Opus 5 | ProviderAnthropic | How you get itHosted | SizeNot published | ReleasedJuly 2026 | StatusSuperseded | What it does | |
Anthropic's model for agentic coding and long enterprise work, sitting below Fable 5.1 and above Sonnet 5. Its thinking adapts to an effort setting rather than a manual budget. It arrived at the same price as the Opus 4.8 it replaced. Anthropic now lists it as legacy, with Opus 5.5 the current Opus model. What it does
Released July 2026 Checked 8 November 2026
|
|||||||
| ModelDeepSeek-V4-Flash | ProviderDeepSeek | How you get itOpen weights | Size304B total | ReleasedJuly 2026 | StatusSuperseded | What it does | |
The small, fast member of DeepSeek's V4 family, under a plain MIT licence with a million-token window. Reasoning effort can be set low, high or max. DeepSeek has since released a V4.1 Flash with native vision. What it does
Released July 2026 Checked 8 November 2026
Where this came from for DeepSeek-V4-Flash (opens in a new tab) |
|||||||
| ModelFLUX 3 | ProviderBlack Forest Labs | How you get itHosted | SizeNot published | ReleasedJuly 2026 | StatusCurrent | What it does | |
One flow model covering image generation and editing, text to video, image to video and video to video with native audio up to twenty seconds. It launched in early access through the API; an open-weight version was announced but not yet available. What it does
Released July 2026 Checked 8 November 2026
|
|||||||
| ModelGPT-5.6 Sol | ProviderOpenAI | How you get itHosted | SizeNot published | ReleasedJuly 2026 | StatusSuperseded | What it does | |
The top of OpenAI's workhorse tier from July to September 2026, and the model behind most paid ChatGPT conversations in that time. It came with two cheaper siblings, Terra as the balanced default and Luna as the fast one, whose prices were cut in July 2026. GPT-6 Sol replaced it at half the price. What it does
Released July 2026 Checked 8 November 2026
|
|||||||
| ModelKimi K3 | ProviderMoonshot AI | How you get itOpen weights | Size2.8T total, 104B active | ReleasedJuly 2026 | StatusCurrent | What it does | |
An open-weight mixture-of-experts model with native image and video understanding, built for long agentic coding and research sessions. It has 896 experts with 16 active per token, and is the first open-weight model in the trillion-parameter class to be published at this scale. What it does
Released July 2026 Checked 8 November 2026
|
|||||||
| ModelMidjourney V8.2 | ProviderMidjourney | How you get itHosted | SizeNot published | ReleasedJuly 2026 | StatusCurrent | What it does | |
Midjourney's default model, reached through its web app and Discord rather than a public API. It generates at 1024 by 1024 by default and doubles that through its upscalers. It is bought as a subscription, so there is no published per-image rate. What it does
Released July 2026 Checked 8 November 2026
Where this came from for Midjourney V8.2 (opens in a new tab) |
|||||||
| ModelClaude Sonnet 5 | ProviderAnthropic | How you get itHosted | SizeNot published | ReleasedJune 2026 | StatusSuperseded | What it does | |
The middle of Anthropic's range and the default on the free and Pro plans, meant as the balance of speed and intelligence. It shares Opus 5's million-token window and adaptive thinking at less than half the price. Claude Sonnet 5.5 replaced it as the current Sonnet model on 28 September 2026. What it does
Released June 2026 Checked 8 November 2026
Where this came from for Claude Sonnet 5 (opens in a new tab) |
|||||||
| ModelMAI-Voice-2 | ProviderMicrosoft | How you get itHosted | SizeNot published | ReleasedJune 2026 | StatusCurrent | What it does | |
A text to speech model covering fifteen languages, with emotion tags controlling tone at the sentence level. It can copy a voice from as little as five seconds of reference audio, but only for voices that have passed a gated consent process. It is built to stay stable over long generations such as audiobooks. What it does
Released June 2026 Checked 8 November 2026
|
|||||||
| ModelKrea 2 | ProviderKrea AI | How you get itHosted | SizeNot published | ReleasedMay 2026 | StatusCurrent | What it does | |
Krea's first foundation image model, built from scratch. It is aimed at style rather than at transcribing a prompt literally, with a style transfer system that takes up to ten reference images and lets each be weighted. It ships in three sizes, and Krea has since published open-weight siblings. What it does
Released May 2026 Checked 8 November 2026
|
|||||||
| ModelClaude Opus 4.7 | ProviderAnthropic | How you get itHosted | SizeNot published | ReleasedApril 2026 | StatusSuperseded | What it does | |
Anthropic's top model of April 2026, for complex agentic coding and enterprise work. It introduced the tokenizer the current Claude models still use, so a million of its tokens holds rather more text than on earlier models. Anthropic lists it as legacy and still available. What it does
Released April 2026 Checked 8 November 2026
Where this came from for Claude Opus 4.7 (opens in a new tab) |
|||||||
| ModelGPT-5.4 Thinking | ProviderOpenAI | How you get itHosted | SizeNot published | ReleasedMarch 2026 | StatusSuperseded | What it does | |
OpenAI's model for professional work from March 2026, folding the Codex line's coding ability into the main series and adding computer use. Thinking is the reasoning variant. Prompts over 272K tokens are charged at double the input rate for the whole session. What it does
Released March 2026 Checked 8 November 2026
Where this came from for GPT-5.4 Thinking (opens in a new tab) |
|||||||
| ModelLTX-2.3 | ProviderLightricks | How you get itOpen weights | Size22B | ReleasedMarch 2026 | StatusSuperseded | What it does | |
A diffusion transformer that generates video and matching audio together in one model, from text or from an image. It came with a desktop editor so the whole thing runs on a single consumer graphics card. LTX-2.5 replaced it in August 2026. What it does
Released March 2026 Checked 8 November 2026
|
|||||||
| ModelSuno v5.5 | ProviderSuno | How you get itHosted | SizeNot published | ReleasedMarch 2026 | StatusCurrent | What it does | |
The music model most general listeners have actually encountered, generating whole songs with vocals from a text description. This version added singing in a subscriber's own verified voice, and custom models trained on a user's own tracks. It is sold by subscription with credits. What it does
Released March 2026 Checked 8 November 2026
|
|||||||
| ModelClaude Sonnet 4.6 | ProviderAnthropic | How you get itHosted | SizeNot published | ReleasedFebruary 2026 | StatusSuperseded | What it does | |
The February 2026 middle of Anthropic's range. It was the last generation to accept the older manual extended thinking mode, which is deprecated on it in favour of adaptive thinking steered by an effort setting. Claude Sonnet 5 replaced it at a lower price. What it does
Released February 2026 Checked 8 November 2026
Where this came from for Claude Sonnet 4.6 (opens in a new tab) |
|||||||
| ModelGemini 3.1 Pro | ProviderGoogle | How you get itHosted | SizeNot published | ReleasedFebruary 2026 | StatusCurrent | What it does | |
Google's top general model, taking text, images, audio, video and whole code repositories and answering in text. It is built for long reasoning and agentic work, with a Deep Think mode for harder problems. It was still labelled a preview in the Gemini API when this was checked. What it does
Released February 2026 Checked 8 November 2026
Where this came from for Gemini 3.1 Pro (opens in a new tab) |
|||||||
| ModelKling Video 3.0 | ProviderKuaishou | How you get itHosted | SizeNot published | ReleasedFebruary 2026 | StatusCurrent | What it does | |
Kuaishou's video model, and the main non-Western entrant most readers will meet. It generates speech in several languages and accents with separate control of each speaker, and handles multi-shot sequences of up to fifteen seconds. What it does
Released February 2026 Checked 8 November 2026
Where this came from for Kling Video 3.0 (opens in a new tab) |
|||||||
| ModelFLUX.2 [klein] 9B | ProviderBlack Forest Labs | How you get itOpen weights | Size9B, with an 8B text embedder | ReleasedJanuary 2026 | StatusCurrent | What it does | |
A rectified flow transformer doing text to image and reference-based editing in one model. It is step distilled to four inference steps and generates in under a second on suitable hardware, needing about 29GB of video memory. The 4B sibling is Apache licensed, but these 9B weights are non-commercial only. What it does
Released January 2026 Checked 8 November 2026
Where this came from for FLUX.2 [klein] 9B (opens in a new tab) |
|||||||
| ModelLTX-2.5 | ProviderLightricks | How you get itOpen weights | Size22B | ReleasedJanuary 2026 | StatusCurrent | What it does | |
An open-weight model generating synchronised video and audio together, from HD up to 4K, meant to be fine-tuned and run on your own hardware. It can produce several shots that keep the same character, lighting and voice across cuts. The licence is free below a revenue threshold, so it is open weights rather than open source. What it does
Released January 2026 Checked 8 November 2026
|
|||||||
| ModelRunway Gen-4.5 | ProviderRunway | How you get itHosted | SizeNot published | ReleasedDecember 2025 | StatusCurrent | What it does | |
Runway's flagship video model, aimed at film and advertising work rather than short social clips. It is sold through credits on the paid plans rather than a per-second rate. Runway is unusually open about its limits, naming object permanence and causal reasoning as things it still gets wrong. What it does
Released December 2025 Checked 8 November 2026
Where this came from for Runway Gen-4.5 (opens in a new tab) |
|||||||
| ModelNano Banana Pro | ProviderGoogle | How you get itHosted | SizeNot published | ReleasedNovember 2025 | StatusCurrent | What it does | |
Google's image generation and editing model, formally Gemini 3 Pro Image, built on Gemini 3 Pro so that it can reason about what it is drawing. It is the one most often reached for when an image has to carry legible text, such as a poster or an annotated diagram, and it holds several characters consistent across edits. A cheaper Flash tier sits below it. What it does
Released November 2025 Checked 8 November 2026
Where this came from for Nano Banana Pro (opens in a new tab) |
|||||||
| ModelZ-Image-Turbo | ProviderAlibaba | How you get itOpen weights | Size6B | ReleasedNovember 2025 | StatusCurrent | What it does | |
A distilled text to image model that reaches a usable image in about eight steps. It fits in 16GB of video memory, so it runs on consumer hardware, and it renders both English and Chinese text inside the image unusually well. What it does
Released November 2025 Checked 8 November 2026
|
|||||||
| ModelClaude Haiku 4.5 | ProviderAnthropic | How you get itHosted | SizeNot published | ReleasedOctober 2025 | StatusCurrent | What it does | |
The fastest and cheapest model in the Claude range, for high volume and for the sub-agents inside a larger system. It is the only one still on the older 200K window and the manual extended-thinking mode. What it does
Released October 2025 Checked 8 November 2026
Where this came from for Claude Haiku 4.5 (opens in a new tab) |
|||||||
| ModelVeo 3.1 | ProviderGoogle | How you get itHosted | SizeNot published | ReleasedOctober 2025 | StatusCurrent | What it does | |
Google's current video model and the successor to Veo 2. It generates dialogue, sound effects and ambience natively alongside the picture rather than leaving them to be dubbed on, and offers scene extension, object insertion and camera control. Cheaper Fast and Lite tiers cost a fraction of the full rate. What it does
Released October 2025 Checked 8 November 2026
|
|||||||
| ModelGPT-5 | ProviderOpenAI | How you get itHosted | SizeNot published | ReleasedAugust 2025 | StatusSuperseded | What it does | |
OpenAI's August 2025 flagship, built as a router that sends a request either to a fast general model or to a deeper reasoning one, with reasoning effort as a dial. OpenAI's own documentation now marks it superseded. The system card at openai.com refuses to be fetched, so this rests on the developer documentation. What it does
Released August 2025 Checked 8 November 2026
|
|||||||
| ModelQwen-Image | ProviderAlibaba | How you get itOpen weights | Size20B | ReleasedAugust 2025 | StatusCurrent | What it does | |
The Qwen team's first open-weight image model, under a fully permissive licence. Its distinguishing strength is rendering long passages of text inside an image, in Chinese as well as English, which most open models handle badly. It runs locally and has companion editing checkpoints. What it does
Released August 2025 Checked 8 November 2026
|
|||||||
| ModelGrok 4 | ProviderxAI | How you get itHosted | SizeNot published | ReleasedJuly 2025 | StatusRetired | What it does | |
xAI's flagship of mid 2025, with native tool use and live search over X and the web. It was retired from the API on 15 May 2026; the old name still resolves but is served by Grok 4.3 and billed at its rates. What it does
Released July 2025 Checked 8 November 2026
|
|||||||
| ModelWan 2.2 T2V | ProviderAlibaba | How you get itOpen weights | Size27B total, 14B active | ReleasedJuly 2025 | StatusCurrent | What it does | |
The strongest openly licensed text to video model most people can actually download, generating five seconds at up to 720p. It uses a pair of experts, one for the early layout and one for the later detail, so it holds more capacity without costing more to run. Alibaba's later Wan releases moved behind an API, which makes this the last with published weights. What it does
Released July 2025 Checked 8 November 2026
|
|||||||
| ModelGemini 2.5 Flash | ProviderGoogle | How you get itHosted | SizeNot published | ReleasedJune 2025 | StatusCurrent | What it does | |
Google's price-performance model for high volume work that still needs some reasoning. It takes text, images, audio and video, and its thinking budget can be turned down to trade depth for cost. What it does
Released June 2025 Checked 8 November 2026
Where this came from for Gemini 2.5 Flash (opens in a new tab) |
|||||||
| ModelGemini 2.5 Pro | ProviderGoogle | How you get itHosted | SizeNot published | ReleasedJune 2025 | StatusSuperseded | What it does | |
Google's reasoning model in the 2.5 family, which thinks before answering and takes long documents, code and video. It is still served with no announced shutdown, but Google now fronts the Gemini 3 series. What it does
Released June 2025 Checked 8 November 2026
Where this came from for Gemini 2.5 Pro (opens in a new tab) |
|||||||
| ModelGPT-4.1 | ProviderOpenAI | How you get itHosted | SizeNot published | ReleasedApril 2025 | StatusSuperseded | What it does | |
An API-only successor to GPT-4o, described by OpenAI as its strongest non-reasoning model, with better instruction following and tool calling. It raised the window to about a million tokens. It has gone from ChatGPT, though the API model has no announced shutdown. What it does
Released April 2025 Checked 8 November 2026
|
|||||||
| ModelLlama 4 Scout | ProviderMeta | How you get itOpen weights | Size109B total, 17B active | ReleasedApril 2025 | StatusCurrent | What it does | |
The smaller of the two released Llama 4 models, a mixture of experts with about 17 billion parameters active per token. It takes text and images and writes text and code. Its headline feature is a context window of ten million tokens, far beyond anything else here. What it does
Released April 2025 Checked 8 November 2026
|
|||||||
| Modelo3 | ProviderOpenAI | How you get itHosted | SizeNot published | ReleasedApril 2025 | StatusSuperseded | What it does | |
A reasoning model that works through a problem in text, code and images before answering, aimed at maths, science and visual reasoning. Its April 2025 snapshot is deprecated with a shutdown date of 11 December 2026, and OpenAI names a GPT-5.6 model as the replacement. What it does
Released April 2025 Checked 8 November 2026
|
|||||||
| ModelCommand A | ProviderCohere | How you get itOpen weights | Size111B | ReleasedMarch 2025 | StatusSuperseded | What it does | |
Cohere's enterprise model for retrieval, tool use and multilingual work. Unusually for its size it runs on as few as two datacentre cards, which is why organisations pick it for private deployment. It has been overtaken by Command A+. What it does
Released March 2025 Checked 8 November 2026
|
|||||||
| ModelGemma 3 27B | ProviderGoogle | How you get itOpen weights | Size27B | ReleasedMarch 2025 | StatusSuperseded | What it does | |
The largest of the Gemma 3 open models, meant to run on a single card while taking both text and images. It was trained on 14 trillion tokens across more than 140 languages. Gemma 4 arrived in March 2026, so this is no longer the generation to start from. What it does
Released March 2025 Checked 8 November 2026
|
|||||||
| ModelDeepSeek-R1 | ProviderDeepSeek | How you get itOpen weights | Size671B total, 37B active | ReleasedJanuary 2025 | StatusSuperseded | What it does | |
A reasoning model built on the V3 architecture and trained with reinforcement learning, which works through a chain of thought before answering. It was the first widely available open-weight model to match OpenAI's o1 on reasoning, and shipped with six smaller distilled versions. What it does
Released January 2025 Checked 8 November 2026
|
|||||||
| ModelHunyuan3D 2.0 | ProviderTencent | How you get itOpen weights | SizeNot published | ReleasedJanuary 2025 | StatusSuperseded | What it does | |
A two-stage system generating textured 3D assets from an image or a prompt: a shape model makes the geometry and a separate model paints it, so either stage can be used alone. Tencent has since released 2.1 and a 3.0 generation. What it does
Released January 2025 Checked 8 November 2026
|
|||||||
| ModelMistral Small 3 | ProviderMistral | How you get itOpen weights | Size24B | ReleasedJanuary 2025 | StatusSuperseded | What it does | |
A dense 24B model under a permissive licence, tuned for low latency rather than maximum capability, with native function calling. Later revisions moved quickly: 3.1 extended the window to 128K and added vision. What it does
Released January 2025 Checked 8 November 2026
Where this came from for Mistral Small 3 (opens in a new tab) |
|||||||
| ModelEleven v3 | ProviderElevenLabs | How you get itHosted | SizeNot published | Released2025 | StatusCurrent | What it does | |
ElevenLabs' flagship speech synthesis model, covering more than seventy languages. It is the counterpart to Whisper here, turning text into speech rather than speech into text, and is chosen for audiobooks and character dialogue where emotional range matters. A separate conversational variant trades some quality for much lower latency. What it does
Released 2025 Checked 8 November 2026
|
|||||||
| ModelDeepSeek-V3 | ProviderDeepSeek | How you get itOpen weights | Size671B total, 37B active | ReleasedDecember 2024 | StatusSuperseded | What it does | |
A mixture-of-experts chat model pre-trained on 14.8 trillion tokens in FP8. Its interest was cost: near-frontier quality for a fraction of the usual training compute. The original weights carry DeepSeek's own licence, not the MIT that arrived with the March 2025 refresh. What it does
Released December 2024 Checked 8 November 2026
|
|||||||
| ModelNova Pro | ProviderAmazon | How you get itHosted | SizeNot published | ReleasedDecember 2024 | StatusSuperseded | What it does | |
Amazon's mid-tier multimodal model on Bedrock, taking text, images and video. Its model card still shows it active, but Amazon has announced Nova 2 and moved several sibling models to legacy. What it does
Released December 2024 Checked 8 November 2026
|
|||||||
| ModelPhi-4 | ProviderMicrosoft | How you get itOpen weights | Size14B | ReleasedDecember 2024 | StatusCurrent | What it does | |
Microsoft's open-weight small model, trained heavily on synthetic and curated data so that fourteen billion parameters perform above their size on reasoning and maths. It is still the current generation: there is no Phi-5, and the family has grown sideways instead, adding mini, multimodal, reasoning and vision variants. What it does
Released December 2024 Checked 8 November 2026
|
|||||||
| ModelSora | ProviderOpenAI | How you get itHosted | SizeNot published | ReleasedDecember 2024 | StatusRetired | What it does | |
The model that made text to video widely known. It was replaced by Sora 2 in September 2025, the apps closed in April 2026, and the API stops accepting requests on 24 September 2026. What it does
Released December 2024 Checked 8 November 2026
|
|||||||
| ModelVeo 2 | ProviderGoogle | How you get itHosted | SizeNot published | ReleasedDecember 2024 | StatusRetired | What it does | |
Text and image to video at up to 1080p, silent, with control over the first and last frame. Google shut it down on 30 June 2026 and directs users to Veo 3.1. What it does
Released December 2024 Checked 8 November 2026
|
|||||||
| ModelQwen 2.5 Coder 32B | ProviderAlibaba | How you get itOpen weights | Size32.5B | ReleasedNovember 2024 | StatusSuperseded | What it does | |
A code-specialised model trained on about 5.5 trillion tokens of source and synthetic code. At release it was the strongest open-weight coding model of its size and small enough for one machine. Alibaba has since moved to the Qwen3-Coder line. What it does
Released November 2024 Checked 8 November 2026
Where this came from for Qwen 2.5 Coder 32B (opens in a new tab) |
|||||||
| ModelStable Diffusion 3.5 Large | ProviderStability AI | How you get itOpen weights | Size8.1B | ReleasedOctober 2024 | StatusCurrent | What it does | |
A text to image diffusion transformer generating at up to about a megapixel, with better text rendering and prompt adherence than Stable Diffusion 3. The licence is free below a revenue threshold and paid above it, so it is not open source in the usual sense. It remains Stability's current open image release. What it does
Released October 2024 Checked 8 November 2026
Where this came from for Stable Diffusion 3.5 Large (opens in a new tab) |
|||||||
| ModelQwen 2.5 72B | ProviderAlibaba | How you get itOpen weights | Size72.7B | ReleasedSeptember 2024 | StatusSuperseded | What it does | |
The largest of the Qwen2.5 series, covering more than 29 languages and strong at code, mathematics and structured output. It reaches its long window by extrapolation rather than native training. Unlike the smaller Qwen2.5 models it is not Apache-licensed. What it does
Released September 2024 Checked 8 November 2026
|
|||||||
| ModelFLUX.1 [dev] | ProviderBlack Forest Labs | How you get itOpen weights | Size12B | ReleasedAugust 2024 | StatusSuperseded | What it does | |
A rectified flow transformer distilled from the hosted model, so it runs in fewer steps. At release it was the largest openly available text to image model, and notably better at text inside an image than what came before. The licence forbids commercial and production use. What it does
Released August 2024 Checked 8 November 2026
|
|||||||
| ModelImagen 3 | ProviderGoogle | How you get itHosted | SizeNot published | ReleasedAugust 2024 | StatusRetired | What it does | |
Google DeepMind's text to image model on Vertex AI, with standard and fast tiers, editing and upscaling, and SynthID watermarking. Google has deprecated the Imagen endpoints and routes image work to its Gemini-native models. What it does
Released August 2024 Checked 8 November 2026
|
|||||||
| ModelLlama 3.1 405B | ProviderMeta | How you get itOpen weights | Size405B | ReleasedJuly 2024 | StatusSuperseded | What it does | |
The first openly released model at frontier scale, trained on about 15 trillion tokens across eight languages. Running it needs several hundred gigabytes of accelerator memory, so most people reach it through a host rather than their own hardware. What it does
Released July 2024 Checked 8 November 2026
Where this came from for Llama 3.1 405B (opens in a new tab) |
|||||||
| ModelLlama 3.1 70B | ProviderMeta | How you get itOpen weights | Size70B | ReleasedJuly 2024 | StatusSuperseded | What it does | |
The middle size of the Llama 3.1 family, which fits on a single high-memory node when quantised. Meta's own Llama 3.3 70B matches much of the 405B model's quality at this size and has largely displaced it. What it does
Released July 2024 Checked 8 November 2026
|
|||||||
| ModelLlama 3.1 8B | ProviderMeta | How you get itOpen weights | Size8B | ReleasedJuly 2024 | StatusSuperseded | What it does | |
The smallest Llama 3.1, for local and on-device use and for fine-tuning on narrow tasks. It runs on a consumer graphics card when quantised. Its long life as the default local model owes more to its size than its quality, which later small models have passed. What it does
Released July 2024 Checked 8 November 2026
|
|||||||
| ModelGemma 2 27B | ProviderGoogle | How you get itOpen weights | Size27B | ReleasedJune 2024 | StatusSuperseded | What it does | |
A text-only open model built on the research behind Gemini, sized to run on a single desktop or a modest cloud instance. Its 8K window is short by current standards, and Google has since released Gemma 3 and 4. What it does
Released June 2024 Checked 8 November 2026
|
|||||||
| ModelGPT-4o | ProviderOpenAI | How you get itHosted | SizeNot published | ReleasedMay 2024 | StatusSuperseded | What it does | |
The model that made speech in and speech out ordinary, and brought native multimodal handling to the GPT-4 line. It was withdrawn from ChatGPT in April 2026, but the API snapshots are still served with no announced shutdown. What it does
Released May 2024 Checked 8 November 2026
|
|||||||
| ModelClaude 3 Opus | ProviderAnthropic | How you get itHosted | SizeNot published | ReleasedMarch 2024 | StatusRetired | What it does | |
The largest of the three Claude 3 models and the first Anthropic model to read images as well as text. Its long window and its writing made it the model many people reached for when a task needed long documents held in mind at once. It was retired in January 2026, the first model to go through Anthropic's full preservation and retirement process. What it does
Released March 2024 Checked 8 November 2026
|
|||||||
| Modelnomic-embed-text-v1.5 | ProviderNomic AI | How you get itOpen weights | Size0.1B | ReleasedFebruary 2024 | StatusCurrent | What it does | |
A small embedding model for retrieval and semantic search, with a long window for its size. Its vectors can be truncated from 768 dimensions to as few as 64 with only a small loss of accuracy. Queries and documents must carry a task prefix. What it does
Released February 2024 Checked 8 November 2026
Where this came from for nomic-embed-text-v1.5 (opens in a new tab) |
|||||||
| ModelMixtral 8x7B | ProviderMistral | How you get itOpen weights | Size46.7B total, 12.9B active | ReleasedDecember 2023 | StatusSuperseded | What it does | |
A sparse mixture of experts in which a router picks two of eight expert groups per token, so it costs about as much to run as a 13B model while holding far more. It made open-weight mixture-of-experts designs ordinary. What it does
Released December 2023 Checked 8 November 2026
|
|||||||
| ModelStable Video Diffusion | ProviderStability AI | How you get itOpen weights | Size2B | ReleasedNovember 2023 | StatusSuperseded | What it does | |
A latent diffusion model turning a single still into a short clip, 25 frames at 576 by 1024 in the XT variant. It was released as a research preview and was one of the first open-weight video models of any quality. Stability has withdrawn it from its own API. What it does
Released November 2023 Checked 8 November 2026
Where this came from for Stable Video Diffusion (opens in a new tab) |
|||||||
| ModelWhisper large-v3 | ProviderOpenAI | How you get itOpen weights | Size1.55B | ReleasedNovember 2023 | StatusSuperseded | What it does | |
Speech recognition and speech-to-English translation across 99 languages, trained on more than five million hours of audio. It takes 30 seconds at a time, so longer recordings are chunked. OpenAI's hosted endpoint is being removed in February 2027, but the open weights remain the last major Whisper release. What it does
Released November 2023 Checked 8 November 2026
Where this came from for Whisper large-v3 (opens in a new tab) |
|||||||
| ModelDALL·E 3 | ProviderOpenAI | How you get itHosted | SizeNot published | ReleasedOctober 2023 | StatusRetired | What it does | |
The image model that followed long prompts closely and rewrote a short request before generating. OpenAI removed it from the API on 12 May 2026 and points developers at its GPT Image line instead. What it does
Released October 2023 Checked 8 November 2026
|
|||||||
| ModelLlama 2 70B | ProviderMeta | How you get itOpen weights | Size70B | ReleasedJuly 2023 | StatusSuperseded | What it does | |
The largest of Meta's second Llama release, pretrained on two trillion tokens and offered as a base model and a chat-tuned one. Unlike the first Llama it came with a licence permitting commercial use, which is the moment open-weight models became a serious alternative to hosted APIs for ordinary work. What it does
Released July 2023 Checked 8 November 2026
|
|||||||
| ModelMusicGen | ProviderMeta | How you get itOpen weights | Size300M, 1.5B and 3.3B | ReleasedJune 2023 | StatusCurrent | What it does | |
A text to music model that writes short instrumental pieces from a written description, optionally following a melody you give it. It was trained on 20,000 hours of licensed music with vocals stripped out. It cannot sing, and it was trained on English descriptions only. What it does
Released June 2023 Checked 8 November 2026
|
|||||||
| ModelPaLM 2 | ProviderGoogle | How you get itHosted | SizeNot published | ReleasedMay 2023 | StatusRetired | What it does | |
The model behind Bard and Google's first generation of hosted language APIs, sold in four sizes named Gecko, Otter, Bison and Unicorn. Its technical report withheld the parameter count, the dataset size and the context length, which drew criticism at the time. Google replaced it with Gemini and shut the API down in 2024. What it does
Released May 2023 Checked 8 November 2026
|
|||||||
| ModelGPT-4 | ProviderOpenAI | How you get itHosted | SizeNot published | ReleasedMarch 2023 | StatusSuperseded | What it does | |
The model that made language models ordinary office equipment, passing professional exams and accepting images as well as text. Its system card described the safety testing but withheld the architecture, parameter count and training data, which marked the point where frontier labs stopped describing their models. Its last snapshot shuts down on 23 October 2026. What it does
Released March 2023 Checked 8 November 2026
|
|||||||
| ModelGALACTICA | ProviderMeta | How you get itOpen weights | Size120B | ReleasedNovember 2022 | StatusRetired | What it does | |
Trained on 106 billion tokens of open-access scientific text for citation prediction, scientific question answering and molecular property prediction. Meta withdrew the public demo after three days because the model produced confident, fluent and entirely invented scientific claims. It is the standard example of why fluency in a specialist register is not the same as being right. What it does
Released November 2022 Checked 8 November 2026
|
|||||||
| ModelStable Diffusion 1.5 | ProviderCompVis and Runway | How you get itOpen weights | SizeAbout 0.9B | ReleasedOctober 2022 | StatusSuperseded | What it does | |
A latent text to image model trained on the LAION-2B English dataset with a frozen CLIP text encoder. Because the weights were published and small enough for a consumer graphics card, it became the base that thousands of fine-tunes and local image tools were built on. Its training data also made it the focus of the first copyright disputes over image models. What it does
Released October 2022 Checked 8 November 2026
Where this came from for Stable Diffusion 1.5 (opens in a new tab) |
|||||||
| ModelOPT-175B | ProviderMeta | How you get itOpen weights | Size175B | ReleasedMay 2022 | StatusSuperseded | What it does | |
Meta's open replication of GPT-3, released to researchers with the training code and a logbook of what went wrong during the run. That logbook was as valuable as the weights: it was the first candid public account of what training a model at that scale actually involves. What it does
Released May 2022 Checked 8 November 2026
|
|||||||
| ModelDALL·E 2 | ProviderOpenAI | How you get itHosted | SizeNot published | ReleasedApril 2022 | StatusRetired | What it does | |
Generated images from text and could edit a region of an uploaded picture or produce variations of it. It opened to roughly 400 vetted users, and the wait list and the quality of the results together started the public excitement about generative images. It left the API in May 2026. What it does
Released April 2022 Checked 8 November 2026
|
|||||||
| ModelCLIP | ProviderOpenAI | How you get itOpen weights | SizeNot published | ReleasedJanuary 2021 | StatusSuperseded | What it does | |
A model that learned to put images and their captions in one shared space, so it could classify pictures against categories it had never been trained on, described in plain words. Its text encoder went on to steer most of the first generation of text to image models, Stable Diffusion among them. What it does
Released January 2021 Checked 8 November 2026
|
|||||||
| ModelGPT-3 | ProviderOpenAI | How you get itHosted | Size175B | ReleasedJune 2020 | StatusRetired | What it does | |
The first very large language model sold as a service, reachable only through an API under controlled access. Its demonstration that a model could learn a task from a few examples in the prompt, with no fine-tuning, is what made prompting a discipline. The base models were shut down in January 2024. What it does
Released June 2020 Checked 8 November 2026
|
|||||||
| ModelGPT-2 | ProviderOpenAI | How you get itOpen weights | Size1.5B | ReleasedFebruary 2019 | StatusSuperseded | What it does | |
A text completion model trained on a corpus drawn from 45 million links posted to Reddit. OpenAI released it in stages through 2019 because it feared the largest version would be misused, and that decision set off the first serious public argument about whether weights should be published. What it does
Released February 2019 Checked 8 November 2026
|
|||||||
No model matches those filters.