MiniMax, Hunyuan, StepFun and Ling: the model makers you have not heard of
Four lesser-known AI labs, two still in the MultiChats picker and two that left it in October. What each one is for, and an honest verdict on each.
Scroll far enough down the model list and the names stop being familiar. MiniMax M3. Ling 3.0 Flash. Until early October 2026, Hunyuan 3 and Step 3.7 Flash sat there too. An unfamiliar name gives you little reason to choose it.
That instinct is mostly correct. A model you have never heard of is not owed a turn just for existing, and "it exists" is not a reason to spend an afternoon on it. But one of these four makers is worth a real try, one is on the free plan, and two have already left the picker, which is worth knowing if you meet their names in an older chat.
So: four makers, the same five questions asked of each one still in the picker, a short account of the two that left, and a verdict at the end of every entry, including the entries where the verdict is no. (Qwen, GLM, Kimi and Mistral also sit in the picker and are covered separately.)
What the three models still here have in common
Before the entries, the ground they all share, so the entries do not have to repeat it.
They are all reasoning models, and on all three the thinking cannot be switched off. Only a minority of the reasoning models in the picker offer an Off setting at all, and none of these three does, so every answer carries the pause. Our explainer on reasoning models covers when that pause earns its keep.
None of them accepts a PDF attachment. The paperclip will refuse the file rather than send it. Only one of them takes images: MiniMax M3.
None of them runs a web search of its own. Whatever their makers claim on that front, the answer here comes from what the model knows plus what you put in front of it, unless you have pointed the search step at another model in your settings.
None of these three is served to you by the company that made it. We route them through third-party infrastructure and restrict which providers are eligible, and a test in our build fails if a model from one of these makers is added without that restriction in place. Each model's info card in the picker carries the rest of its capability detail.
And the part worth stating plainly rather than hedging: we do not train on your conversations. What the provider behind a given model does with them varies by model, and the models where we know of a difference carry a notice on their info card. None of these three does, so the card is not where you settle this question.
MiniMax
Who they are. Founded in early 2022, according to its own about page, which lists a row of consumer products alongside the models: MiniMax Code, MiniMax Design, MiniMax Audio and Talkie. Of the four makers here, MiniMax is the one most likely to have reached you already under a product name rather than a company name.
What you get here. Two models, both on Pro. MiniMax M3 is configured at a 524,288-token window and accepts image uploads. MiniMax M2.7 is configured at 196,608 tokens and is text only. Both think, and neither lets you stop it.
Their own pitch. MiniMax bills M3 on its site as a frontier coding and agentic model with a novel attention architecture and a million tokens of context. The Hugging Face model card puts it at roughly 428B total parameters with about 23B active, trained across text, image and video from the first step rather than bolted together afterwards. The same card credits its sparse-attention design with 9x prefill and 15x decode speedups over the earlier M2 at 1M context.
Where it stops. The million-token figure is MiniMax's; the window you get in this app is 524,288, which is still the largest window of the models covered here, though it falls short of the marketing figure. M2.7 has no capability we expose that M3 lacks: smaller window, no images, older generation. It is in the list for continuity, not because it wins at anything.
Verdict: the strongest of the models still here on paper and the one to try first, on a long pasted document with a screenshot attached. Skip M2.7 unless you specifically want the older model's writing style.
Tencent Hunyuan
Who they are. Tencent, founded in 1998 and headquartered in Shenzhen, per its own corporate site. The same page walks through the business groups behind Weixin, its games and eSports arm, and its cloud division. This is by far the largest company on the list, and the one whose AI work is least visible outside China.
What happened here. Hunyuan 3, which Tencent calls Hy3, left the picker on 9 October 2026. None of the providers we allow for models from these makers served it any more, so we switched it off. A chat that used it now continues on Qwen3.6 35B.
Verdict: nothing to open here for now. If an older chat of yours ran on Hunyuan 3, read its later replies as coming from a different model.
StepFun
Who they are. A lab that goes by 阶跃星辰 at home and StepFun in English, with the thinnest English-language footprint of the four. Its own site sells an API platform and a build tool called AI Studio, and now leads on Step 5 Preview, a larger flagship announced on 1 October 2026 whose weights its homepage now lists as available. Step 3.7 Flash was the one we carried.
What happened here. Step 3.7 Flash left the picker at the start of October 2026, for the same reason: none of the providers we allow served it any more. A chat that used it now continues on DeepSeek V4.1 Flash. We do not carry Step 5 Preview.
Verdict: nothing to open here for now. If you liked handing Step 3.7 Flash a screenshot or a chart, MiniMax M3 is the model in this group that still takes images.
InclusionAI Ling
Who they are. InclusionAI describes itself on its Hugging Face organisation page as "home for Ant Group's Artificial General Intelligence (AGI) initiative". Ant Group's own homepage says the company "traces its roots back to Alipay, which was established in 2004". The same page calls Ant Group "one of the world's leading open Internet platforms" and leads on digital payments. So the least familiar name on this list has the most familiar business behind it.
What you get here. One model, Ling 3.0 Flash, and it is the only one of the four available on the free plan. A 131,072-token window, text only, reasoning always on; a free account cannot set the effort, and the app caps the level it picks itself at medium.
Their own pitch. The model card describes 124B total parameters with 5.1B activated per token, a hybrid linear attention design used from the start of pretraining rather than retrofitted, 512 routed experts with 8 active, and a context schedule trained out to 256K. The claim it leads with is speed under load: a reduction in time to first token of 60% to over 80% on long inputs. It is MIT licensed.
Where it stops. Smallest window of the models still here, at half what the card trains to. No images, no PDFs. One number to ignore if you meet it elsewhere: several write-ups credit Ling 3.0 Flash with 51B active parameters. The model card says 5.1B, which is a tenth of that and the whole point of the design.
Verdict: the one of these four most readers will actually meet, because it is the only one a free account can open. Give it a long reasoning question typed as text and see whether the answer comes back faster than you expected.
If you try exactly one
Try MiniMax M3, on something long, with an image attached. It has the widest window of the three still in this app, it is the only one that takes both a long document and an image, and if it disappoints you, the other two are unlikely to change your mind.
If you are on the free plan, the choice is made for you: Ling 3.0 Flash is the only one of these you can open, and that is a fine place to find out whether an unfamiliar name is worth your attention at all.
And if the answer to all of this is that you would rather stay with the models you know, there is no need to switch. Four entries, four verdicts, and only one of them is an enthusiastic yes.