Is Google Delaying Gemini 4 Pro? Kimi K3.1 and MiniMax M3.1 LEAKED!

Author: Universe of AI · Watch on YouTube

Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.

Original language: English

Original video: Is Google Delaying Gemini 4 Pro? Kimi K3.1 and MiniMax M3.1 LEAKED! · Channel: Universe of AI · 09:20 · Watch on YouTube

Topics: AI evaluation and benchmarks, AI models, Multimodal models

Summary

The video reports that Google has announced a new Gemini 3.8 Flash text-to-speech model, while there is still no confirmed release date for Gemini 4 Pro. The host discusses an X post that appears to label an output as Gemini 4 Pro, but says it is unclear whether this indicates testing or an upcoming launch. The video then examines Space Bunny, which the host believes is likely a stealth test of MiniMax M3.1, while cautioning that early outputs may reflect harness issues. Finally, it covers signs of an upcoming Kimi K3.1 release and the host’s expectations for the model.

Key points

Google’s newly announced Gemini 3.8 Flash is a text-to-speech model; the video does not provide a confirmed launch date for Gemini 4 Pro.
An X user reportedly saw an output labeled “created with 4 Pro,” but the host says this is not enough to confirm a launch or stealth testing.
The host argues that delays to Gemini 4 Pro give competing labs more time to release new models.
Space Bunny is described as a multimodal model with a one-million-token context window and zero data retention, and the host says it is likely a MiniMax model.
Early Space Bunny tests, including a Minecraft-clone task and 3D-generation prompts, appear weak to the host, who suggests a testing-harness issue could be responsible.
The host cites repeated error behavior and a MiniMax catalog reference as evidence that Space Bunny is likely MiniMax M3.1.
The host says Space Bunny can be tried through platforms including OpenCode and OpenRouter.
Kimi K3.1 has reportedly appeared in Moonshot-related posts and source code; the host speculates it could arrive that week or early the next week.
The host says early reports suggest Kimi K3.1 may match Opus 5, but doubts it will match the performance of Opus 5.5 or GPT 5.6.

Chapters

Start	Chapter	Summary
00:00	Gemini 3.8 Flash and the Gemini 4 Pro question	The host introduces Google’s new text-to-speech model, notes the lack of Gemini 4 Pro news, and discusses an ambiguous output label that may reference it.
03:07	Space Bunny and possible MiniMax M3.1 tests	The host presents Space Bunny as a likely MiniMax model and reviews early coding and 3D-generation tests, while allowing that the testing setup may be at fault.
07:46	Signs of an upcoming Kimi K3.1	The host points to Moonshot-related clues and source-code references as possible indications that Kimi K3.1 will be released soon.
08:49	Expectations for Kimi K3.1 and sign-off	The host tempers speculation that Kimi K3.1 could outperform Opus 5.5 and closes with channel and newsletter reminders.

Quotes

Today we got a new model from Google and guess what? It is not Gemini 4 Pro. — 00:00

Does this mean that Google is getting ready to drop the model soon? Is it being stealth tested? I don’t know. — 02:04

I think this is just probably a harness issue rather than an actual model issue. — 04:41

I think it’s already Wednesday, so that leaves us Thursday and Friday. — 08:17

Concepts and entities

Gemini 4 Pro, Gemini 3.8 Flash TTS, Space Bunny, MiniMax M3.1, Kimi K3.1, Opus 5.5, Google DeepMind, MiniMax, Moonshot AI, OpenCode, Harness

Full transcript (original language)

00:00 Today we got a new model from Google and guess what? It is not Gemini 4 Pro. And not only that, it seems like MiniMax and also the Moonshot team are getting ready to drop new models. So, let’s get into it. Today we got a new model from the Google DeepMind team and guess what guys? It is Gemini 3.8 flash text-to-speech. Yes, this is their new state-of-the-art text-to-speech model that has a new design experience, 2,000 plus production-ready voices, voice replication, and support for 100 languages. Voice remixing is also coming

00:32 soon and it’s the number one model on Hume’s AI’s voice benchmark. And I’m sure that you guys are probably excited to know that we have a new text-to-speech model from the Google DeepMind team. That was a joke and I hope you guys caught that on because it looks like we still have no information about when Gemini 4 Pro is coming. Although there’s one weird leak that I found today thanks to one person on X that I’m going to show you guys where something was generated by Gemini 4 Pro and the model actually spits that out. But before I do

01:02 that, I just want to talk about couple of things that I think it’s going to be a problem for the Google DeepMind team cuz the more and more they delay Gemini 4 Pro, they’re giving more opportunity to OpenAI to drop newer models. And we saw that with Opus 5.5. That model is a completely different level of model and if Gemini 4 continues to get delayed, the competition that it’s going to be competing with is just going to get more tough for them because we saw that with many of the newer models. We’re also seeing couple of leaks today that even open-source labs like the MiniMax team

01:34 as well the Moonshot AI team are getting ready to drop newer models. So yeah, I do not know where Gemini 4 Pro is but I’m going to show you guys a output that I saw on X and I want to give credit to the person that actually posted this was this individual. He created something, I guess a friendly greeting or whatever. He just said hello to the model and the model spit out something called created with 4 Pro and this is the timestamp, which was on September 23rd at 6:00 a.m. And he just said, “Hello, how can I help you today?” Which is interesting. I

02:04 don’t know if I have seen this before where the Gemini output is saying that it’s created with 4 Pro. You guys test this out yourself and like, you know, let’s see if you guys get this output as well. I think when you share your chat history, when he was trying to do that, he tested this out and it was showing that it said Gemini 4 Pro. But you guys test it out yourself. Does this mean that Google is getting ready to drop the model soon? Is it being stealth tested? I don’t know. This is the first time where I’m seeing something like this, so I don’t know what the case is going to be, but they need to drop something soon

02:36 because more and more labs are getting ready to drop something new this week or probably in a couple of weeks, and I have higher confidence that it comes out before Gemini 4 Pro does. Hey guys, thank you once again for all the support on YouTube. The channel has been growing thanks to you guys. And if you haven’t subscribed yet, just hit that subscribe button. It’s a free way to support me and helps me continue making amazing videos. Now, let’s get back into the video because today what we have started to see is that on open code, there’s a model called Space Bunny, and it’s Space Bunny, nothing else, so don’t confuse it

03:07 with anything else, is going to be stealth tested, and this model has 1 million contacts, it’s multimodal, and there’s zero data retention. Now, a lot of people are testing this model out and and I think a lot of people were able to figure out what this model was because some people asked it in Chinese what the model was and the model kind of spit out that it is a MiniMax model. For example, let me show you. We need answers in Chinese user asked, “What model are you based on?” And then it says that it is the MiniMax model, and this is most likely going to be the next MiniMax M3.1

03:37 model. A lot of people haven’t done a lot of testing with this yet, but I’m going to show you guys couple of outputs that it’s generating. And to be honest, it doesn’t look like a crazy jump from MiniMax M3. This model obviously probably is going to be a point one of a jump because the outputs are nothing that you are going to, you know, be really shocked for because these outputs are quite generic. Sometimes it doesn’t do a good job of even following the prompt, but I’m just going to show you the output so you guys can be the judge. Before we continue, we just launched the Universe of AI newsletter. If you want

04:09 to stay on top of AI news without having to hunt for it, link is in the description. Don’t miss out. For example, this model is being compared with DeepSeek version 4.1 flash at the top and then at the bottom you see Space Bunny Alpha, the model that is going to be the MiniMax M3 model, and this is the famous build a Minecraft clone in HTML test. And when it was told to do that, you can see the output does not look that great. The world does not seem fully generated. There are some obvious glitches, and I think this is performing worse than MiniMax M3, but keep in mind

04:41 this is being tested on open code, so it could just could be a harness issue, but when the model actually comes out on MiniMax’s own platform, it performs better than this. I hope so because this is obviously not something that you would expect in 2026. In my opinion, this is performing worse than MiniMax M3, as I said, and this is not going to be the final state of the model. So, don’t lose hope yet. I think this is just probably a harness issue rather than an actual model issue. But as I said, I don’t expect this model to like, you know, shock everyone. I

05:12 have higher expectations from the upcoming DeepSeek model and the Chimney model rather than the MiniMax model at least at the moment because these tests do not give me any sort of confidence. For example, this is an output that we are seeing where the model was told to create a 3D design, and you can see that the world does not look fully generated. The trees are, I guess, half generated because I think it put the infrastructure in place to generate the surrounding bushes and trees like that, but didn’t fully complete it. And you can see that this person said that this

05:43 is a fast model, created this in 4 minutes, but it originally didn’t really understand the prompt and it created a 3D image instead of a 3D generation in 3.js. So, yeah, this model, I think I’m going to give it the benefit of doubt. I think this is just a harness issue rather than the model. But still, like, you know, I haven’t seen this happen to other models when they’re being stealth tested, so I don’t know what’s really going on. Although this was only launched a couple of hours ago, so maybe it gets fixed on all these platforms. But if you look at it even at the, you

06:14 know, SVG test, this output, if you saw this in 2025, you’re probably like, “Oh my god, this is amazing. The model’s able to do this.” But in 2026, we have definitely seen better outputs. Opus 5.5, I don’t want to compare this against MiniMax M3.1 cuz I don’t think they’re going to, you know, be competitive with each other. I think Opus 5.5 is definitely going to be outperforming MiniMax M3.1. But even if you compare MiniMax M3.1 or Space Bunny, with Kimik3, for example, or even, you

06:45 know, any of the deep seek models at the moment, those models are producing better outputs than you see on screen. So yeah, it doesn’t look super promising at the moment. And just so you guys are aware, with there is high confidence on this being the MiniMax model, because for example, it had the same error from word to word as this person shows. And then every clean run trips MiniMax’s token to her nine, and this is the same thing as the M3 model. You can see in this image, Bunny throws MiniMax’s own 2013 error word for word, only two of 12

07:16 M3 hosts do, so you can see that test over here. And then every clean Bunny run trips a MiniMax glitch token to her nine, which is the same thing in production with the M3 model. And then this is MiniMax’s own catalog M3.1 stage, not white listed yet. So yeah, there’s high confidence that Space Bunny is going to be the MiniMax M3.1 model. So if you guys want to try this out, you have a couple of options. It’s hosted on open code, open router, go try it out. You might have to make an account or so, or some areas let you, you know, play around with the model, run some basic

07:46 tests. So, try it out for yourself. At the same time, while all this is going on, Kimi K3.1 has been spotted once again. We saw this kind of being, not leaked, I would say, but like on a Chinese foreign platform, the Moonshot Lab posted something, which was the pi digits, the point after 3.1. So, and you know, that’s them probably indicating that they’re going to drop 3.1 as well. And you know, we saw that with the post. Now, we’re seeing that in actual production, where we’re seeing K3 D1 agent in like this code. And this

08:17 probably means that we might get I’m just I’m saying we might get Kimi K3.1 this week. I think it’s already Wednesday, so that leaves us Thursday and Friday. I never really see a Chinese lab or any of the labs drop anything on the weekends anyway, so it’s either this week or early next week, because if we’re seeing this come up in like obviously the source code, then you know that it’s going to come out soon. So, I have higher hope for the Kimi K3.1 model, and early leaks are saying that this is supposed to match Opus 5. I don’t think it’s going to match Opus 5.5

08:49 and GPT 5.6 sole level performance. Now, I don’t know if this is going to get close to the GPT 6 sole instead of 5.6 sole, but K3.1 probably is going to drop soon. This person is saying, “Imagine if K3.1 drops and it mugs Opus 5.5.” I don’t think that’s going to happen, but let’s see. I would love for the Moonshot team to kind of, you know, shock everyone. But, that’s it for today’s video. Make sure you guys are subscribed to the channel, follow our new newsletter as well at universeofai.beehive.com,

09:20 as well as subscribe to the main channel World of AI, and support us on X by following the universeofai z as well. Until then, I’ll see you guys in the next video.