AI model names explained
Flash, Mini, Opus, 235B: what each piece of an AI model name signals, and how to read any of our 59 model names.
As of 19 October 2026, MultiChats carries 59 models from 17 makers. Gemini 3.5 Flash Lite. GPT-5.4 Mini. Qwen3 VL 235B. Without knowing the naming conventions, it is hard to tell these entries apart.
They are built to a pattern, and broadly the same one at every company: four parts, in a fairly consistent order. Take a name apart and you know who made the model, which rung of that maker's range it sits on, which generation it belongs to, and what was changed at the end. What a name will never tell you is whether it suits the thing you are about to type, which the closing sections cover.
Four parts, one name
The family. Which line of work the model belongs to, and usually, though not always, who built it.
The size word. Which rung of that maker's own range, in that maker's own vocabulary.
The number. Which generation of the line, counted in the maker's release order.
The suffix. What was changed at the end: a thinking pass, a speciality, sometimes only a code name.
Few names carry all four: Mercury 2 has a family and a number and stops.
MultiChats puts the leading AI models in one app. Try it free
The family, which is not always the maker
A family name groups everything a maker has built on one line of work: Claude, GPT, Gemini, Grok, Qwen, Kimi, GLM, Llama, Nova, Nemotron. Plenty carry no trace of who is behind them. Nova 2 Lite is Amazon's; Inkling Small comes from Thinking Machines, Muse Glimmer 30B from Meta. So the web picker prints the maker under the model name, and both group the catalogue by maker.
Size words belong to whoever coined them
People try to compare these across companies, and the comparison does not survive the trip. A size word ranks models inside one maker's range and says nothing about anybody else's. What each maker says about its own words, from its documentation checked in October 2026:
Opus, Sonnet and Haiku are three of Anthropic's. Its models overview lines them up by comparative latency: Opus moderate, Sonnet fast, Haiku fastest.
Flash and Flash Lite are Google's. Its Gemini API list describes 3.6 Flash as balancing speed with multimodal work on everyday tasks, and Flash Lite as the fastest in its generation, for high-throughput work.
Mini is OpenAI's word for a small model aimed at focused tasks, per its GPT-4o Mini page. We carry GPT-5.4 Mini and GPT-5.4 Nano on one generation, and Nano reads as the smaller of the two.
Lite, Small, Large, Super and Ultra turn up wherever a maker prefers plain English: Nova 2 Lite, Mistral Small 4, Mistral Large 3, Nemotron 3 Ultra.
Max marks the top of a range. Alibaba's model page calls Qwen3.8-Max a flagship.
Pro is Google's other Gemini word; Gemini 3.1 Pro is on its current model list.
Pro and Max do double duty. Both also name subscription plans, which is where the confusion starts. Anthropic's pricing page, checked in October 2026, lists Pro, Max 5x and Max 20x, so the word does one job on a price list and another inside a model name. We carry no Gemini Pro chat model: our Gemini line is Flash and Flash Lite.
What the number counts
The number records where a model falls in its own line's release order. It is not a date, and it carries no size information. Gemini 3.8 Flash and GPT-5.6 Terra are two makers counting their own releases, so 3.8 against 5.6 means nothing.
Each counts in its own style. Anthropic runs Sonnet 4.5, then 4.6, then 5, then 5.5. Moonshot writes a letter in: Kimi K2.6, then K3. Alibaba glues the number to the family, as Qwen3.5. ChatGPT Latest drops it entirely, pointing at whatever OpenAI currently ships under that alias, so the model behind that entry can change while the name stays put.
The tail end of the name
The suffix records what was changed, and it repays reading closely.
Thinking and Reasoning mark a variant that works through a problem before answering. Grok 4.3 and Grok 4.3 Reasoning are two entries in one generation, and only the second shows the Reasoning capability on its card. Kimi K2 Thinking uses the other word for it.
Instruct is the counterpart. Alibaba's card for Qwen3-235B-A22B-Instruct-2507 calls that build the model's non-thinking mode, and says it generates no thinking blocks.
Code narrows a model to one job, as in Kimi K2.7 Code, a model we no longer carry. OSS says the maker published its weights, as GPT-OSS 120B does.
A code name can replace the size word outright. OpenAI's recent models reach our picker as GPT-6.1 Sol, GPT-6 Luna and GPT-5.6 Terra, and its own model list pitches Luna at focused, high-volume tasks. There is nothing in the three words to decode.
The number with a B on the end
A parameter count is the one piece of a name that is a measurement: Qwen3 235B, Qwen3.5 397B, Ministral 3 3B. It counts the trained values inside the model, and it reaches the name because the maker published the weights, which makes the figure public. That Qwen card is precise: 235 billion in total, 22 billion active for any given step. Two very different numbers in one name.
A parameter count is a spec. It records how the model was built and carries no verdict on how it will handle your email. No threshold exists above which a model starts being good at your work.
Three names, taken apart
Gemini 3.5 Flash Lite. Gemini family; generation 3.5; Flash Lite, the rung Google calls its fastest there.
GPT-5.4 Mini. GPT family; generation 5.4; Mini, the small rung. The sibling with no size word, GPT-5.4, is the full-size entry.
Qwen3 VL 235B. Qwen family with generation 3 glued on; VL, Alibaba's shorthand; 235B, the parameter count.
Why a bigger rung is not automatically the right one
A bigger rung is trained for harder work, and much of what people do in a chat window is not harder work: drafting a reply, tidying a paragraph, lifting three dates out of a long email. The size word also says nothing about how much of your conversation a model keeps in view, which is a separate number with its own behaviour, and whether a larger or newer entry suits today's job deserves a check of its own, which we cover separately.
Where the app spells it out
The card behind a name is the specification. On the web the picker opens from the composer; its search box matches the display name, the maker and the underlying model id, so typing instruct, 235b or nano finds entries whose names hide those pieces. Hover the small (i) on a card and the info card opens beside it, with the maker under the model name and one icon per capability the model has: Vision, PDF, Video, Web search, Reasoning, Image gen, Voice. Only what is present gets drawn.
On the phone the catalogue slides up full screen, with the same search box and maker tabs, and each row carries those icons apart from Voice, so check that one on the web card. One chip there has nothing to do with the name: New marks the day a model arrived in MultiChats rather than the day its maker released it.
One last step, when the name has gone
If the name you came for is missing, look for the Deprecated badge before assuming it has gone: a retiring model keeps that badge while it is still listed, and on the web its tooltip names a recommended replacement. What the rest of that sequence does to threads you already started we handle on its own.
The next unfamiliar name
The same approach works when a maker releases another model. Something unfamiliar turns up, you split it into family, size word, generation and suffix, and you know roughly what you are holding before reading the announcement. Then you open its card, because the name was never going to tell you whether it reads PDFs.