Qwen, DeepSeek, Kimi and GLM, explained for normal people
Four unfamiliar model names, one ordinary job, and a straight answer on which to open first, which to skip, and where all four stop.
Qwen. DeepSeek. Kimi. GLM. Headlines often mention these names without explaining who makes the models or what they do. Who builds each one is the easy part, and we have a piece on every maker here. The harder question, and the one worth ten minutes, is whether any of the four deserves the next thing you were about to type into a chat window.
Between them these four makers account for 14 of the 59 models in MultiChats as of 19 October 2026, so nearly a quarter of the app. Five of those 14 sit on the free plan, out of 22 free models in total. That is enough of the catalogue to deserve a decision.
One job, carried all the way through
A model only makes sense against a specific task, so here is one, and every judgement below is made against it. You chaired the residents' association meeting last Tuesday, recorded it on your phone and had it transcribed, and you now have roughly 14,000 words of eight people talking over each other about a roof repair, next year's budget, and whether to change the bin collection contract.
What you need out of that is much narrower than the recording: the decisions as distinct from the discussion, every action someone took on with a date if anyone named one, and the items raised and then quietly dropped, which are the ones that resurface in six weeks. None of it requires looking anything up, because it is all already in the transcript. That is the job, and every recommendation below is measured against it.
A paragraph on each maker
All four are Chinese labs, which is the only thing the names have in common, and it is why the question of where your text actually goes turns up further down. Otherwise they are four unrelated companies with four different reasons for being in this business.
Qwen is Alibaba's family and the largest presence of the four here: five entries, two of them free. New ones land constantly, for reasons that have more to do with Alibaba's cloud business than with any single model, which is its own article. The detail that matters for the transcript is a quirk of the two free entries, Qwen3 235B and Qwen3 VL 235B. Neither of them reasons. The other 12 entries all run a thinking pass before every answer and not one of them lets you switch it off, so those two Qwen models are the only ones in the group that reply straight away.
DeepSeek has two entries: DeepSeek V4 Flash, free and badged Recommended in our picker, and DeepSeek V4.1 Flash, which joined on 4 October 2026 and sits on Pro. DeepSeek's own announcement describes V4 Flash as the smaller of the two in its generation, 284 billion parameters of which 13 billion fire for any one word, which is the design that lets something that size answer at a decent clip. V4 Flash takes text and text files, no images, while V4.1 Flash also reads images, and both its history and one independent study of what it will and will not discuss are covered separately.
Kimi is Moonshot AI's line, and all three of its entries sit behind a subscription. K3 is the one people mean when they say Kimi, and its entry claims a million tokens of context. That is the ceiling on what it will accept, and how much of it the model uses well is a separate measurement, which is worth understanding before you lean on a figure like that. Two of the three read images, K2 Thinking is text only, and all three think before they answer.
GLM comes from Z.ai, and we carry four: GLM 5.3 Flash and GLM 5.2 on the free plan, GLM 5.3 and GLM 5 on Pro. The GLM 5.3 flagship joined the picker on 4 October 2026. Flash is the one to know, also badged Recommended, and Z.ai introduces it on the model card as "the first natively multimodal model in the GLM-5 series". In our app that shows up as one difference you can use: it is the only GLM entry that will look at a picture.
Two further makers sit lower down the same list, MiniMax and Ant Group's Ling, and we have rounded those up elsewhere, along with Tencent Hunyuan and StepFun, whose models have left the picker.
Pick on the shape of the job rather than on the name. All 14 of these entries are built for a pile of text you already have, plus a precise instruction about what to pull out of it. Not one of them is built for a question whose answer is out in the world today.
Which one gets the transcript
GLM 5.3 Flash, first. It is on the free plan, it will hold the whole transcript, it shows its thinking so you can see how it sorted decided from discussed, and if a photograph of the flipchart turns up later it can look at that too without a model change. One practical note: uploading the transcript as a text file needs a paid plan, so a free account pastes it into the message box instead, which at this length is no hardship.
Then, on your next message in the same thread, switch to DeepSeek V4 Flash and give it the identical instruction. The conversation carries over, so a second opinion costs one message. Two answers that disagree about what was decided teach you more than one answer that sounds certain, and both are open to a free account.
Ask both of them to quote the line each action item came from. Eight people talking over each other is exactly the material where an invented detail reads as plausible, and a quote you can find in the transcript is the simplest check there is. If the thinking pass in front of every answer starts to grate, and on a job this mechanical it might, Qwen3 235B is the entry that will simply produce the list. Read that output with more care, because there is nothing on screen between your question and its conclusions.
Kimi we would leave for another day. None of its three entries is open to a free account, and K3 in particular is built for a pile far larger than one meeting.
All 14 stop at the same three walls
The roofer's quote arrives as a PDF. Not one of the 14 will take it, and the refusal happens in the app before the file goes anywhere. The picker has 24 models out of 59 that accept a PDF, so the quote is a job for a different thread.
Then someone wants to know whether the bin contract terms quoted in the room still hold this year. No model in this group runs a web search of its own, so unless you have nominated a separate search model under Settings, then Tool Preferences, the answer comes from what the model already knew plus what you pasted, which on this year's contract terms is not good enough. That gap is where confident invention comes from, and on screen it looks identical to a correct answer.
Voice mode is the third wall. Not one of the 14 supports it, so talking the thing through on the walk home means a model from elsewhere in the picker.
The part about where your text goes
Two things about that pasted transcript are checkable rather than promised. None of these 14 models is served to you by the company that made it: their weights are published, so the copy answering you runs on infrastructure we chose instead of the maker's own. What that "open weights" label does and does not cover is a subject in its own right, and we treat it as one. Whether Qwen in particular is safe to use gets a fuller answer in its own piece.
Training is the second, and it is worth stating precisely. We do not train on your conversations. What the provider behind any given model does with them varies from model to model, and where we know of a difference, that model's information card in the picker carries a notice. Six of the 59 do, and one of those reads "Routed via OpenRouter to third-party providers that may briefly retain requests; they do not train on your data." None of these 14 entries carries a notice, so a card you open in this part of the catalogue will be silent on the question, and a blank card is not an answer in either direction.
Worth your time, then?
Our judgement, and judgement is all it is: two of these 14 have earned an hour of your attention, and the other 12 are worth knowing about for the day a specific job calls for one. GLM 5.3 Flash and DeepSeek V4 Flash both sit on the free plan with a Recommended badge, and both are good at the one thing this group does well, which is working carefully through text you brought with you. Qwen earns its turn when you want the list without the deliberation in front of it. Kimi earns its turn when the pile is enormous and you are already paying for the app.
The limits are the part most write-ups of these four names leave out. Nothing here will look anything up, nothing here will read your PDF, and nothing here will talk to you, and on an ordinary Tuesday those gaps settle more than any quality difference between the models does. So hand them the transcript, which is a job they are suited to, and go elsewhere the moment the job stops looking like that.