Skip to content
fedi.software

AI models of 2023: who released what

Every notable model release month by month: vendor and its country, open or closed weights, type, context window, size of open models — and which tasks the models of each year lead today.

2023 9 releases · 5 with open weights

Models of 2023 leading each task today

December

  1. Mistral 7B Instruct v0.2 HFMistral AIFranceOpen weightsLLM7.2B
  2. C mixtral-8x7b-instruct-v0.1Mistral AIFranceOpen weightsLLM32K

November

  1. Qwen 72B HFAlibabaChinaOpen weightsLLM72.3B
  2. D dall-e-3OpenAIUnited StatesClosedImage
  3. C GPT 4OpenAIUnited StatesClosedLLM8K
  4. C GPT-4 TurboOpenAIUnited StatesClosedVision (VLM)128K

July

  1. Llama 2 7b hf HFMetaUnited StatesOpen weightsLLM6.7B

March

  1. C GPT-3.5 TurboOpenAIUnited StatesClosedLLM16K

February

  1. pythia-6.9b HFEleutherAIUnited StatesOpen weightsLLM7B

Release dates come from the model catalogues (models.dev, LiteLLM, Hugging Face); models known only by the year of release are not shown. The leaders of a year are its models that rank highest on our leaderboards today.

Updated · Sources: models.dev (MIT), LiteLLM (MIT), LMArena (CC BY 4.0), Hugging Face

Arena (LMArena) scores: Arena Leaderboard Dataset by LMArena, CC BY 4.0. Modified by fedi.software (filtered, re-ranked, aggregated). Not endorsed by LMArena.

A release calendar of AI labs

This chart puts model releases on a single line of time, newest year first and grouped by month. Each line is one model with its vendor, the country where that company is based, whether the weights are open or closed, its type and its context window. For open models you also see the number of parameters; for a mixture-of-experts (MoE) model the second number is the active part used for each token.

Only models from AI labs make it in: those rated on our leaderboards or with weights published on Hugging Face. Resellers, routers and community fine-tunes are left out, so the list stays a record of who actually built what.

Reading the badges

  • Open weights — the files can be downloaded and run on your own machine; which model fits your hardware tells you what that takes.
  • Closed — the model is reachable only through the vendor's app or API.
  • LLM is a text model, Vision (VLM) also reads images, and Reasoning works through a problem step by step before it answers: slower and pricier per reply, but stronger on maths, code and planning. Image, video and audio generators carry their own badges.

Closed vendors do not publish parameter counts, which is why that cell is empty for them rather than a guess.

The leaders of each year

Above every year sits a strip with the models of that year that lead text, coding, web development, agents, image and video tasks today. This is a present-day view, not a reconstruction of the past: a model released in spring may have been overtaken by autumn, and the strip shows how its generation holds up now on the leaderboards.

Filters, editions and limits

You can narrow the view by vendor, country, open weights and type, or search for a name. Each year also has its own page, a yearly edition you can link to. Release dates come from model catalogues — models.dev, LiteLLM and Hugging Face — and mark the day a model appeared there, which may differ from the announcement by a few days. Models whose sources give only a year are not shown. The country is that of the company's headquarters, not of its data centres or its users.