Qwen3.5 Can Now Run On Your PHONE!

So, Alibaba introduced Qwen 3.5 in line with smaller open weight models aimed at strong reasoning, multimodal performance, and much lower deployment cost, signaling how quickly open model quality is compressing. Command AI The the real secret is the very very small version that will should be able to run on your on your phone. But, we'll get into that. Let me share my screen. Um and they do have this Qwen 3.5 plus, which is hosted in that Alibaba Cloud Model Studio, which we'll talk about in a second. Now, I'm always interested in the SWe Bench, and again, it's all real-world usage. You know, we can only go by you you can only get an indication of benchmarks just to see, and when things have been benchmarked, it's hard to even know whether to even believe them. But, it's This is this is Opus. This is um Where are we? GPT-5.2. This is Opus. This is Gemini 3 Pro. So, it's scoring better than Gemini 3 Pro on SWe Bench, which is, you know, good to hear good to see, and um it's a it's a cheap Chinese model. Do you know what I mean? So, again, I think this is good for light projects, people just getting into it, people kind of just wanting a cheaper alternative. Um does vision as well. Like, I really like MiniMax M2.5, but the thing is about MiniMax M2.5, which by the way, they don't compare against cuz it does better. I think they compare against uh uh Kimi K2 somewhere. Uh MiniMax M2.5 is only text. Doesn't do vision. So, from a model multimodal perspective, it's quite a quite a good quite a good little thing. I want to do a video on this because their $10 a month plan supports Kimi K2 and MiniMax and GLM and Qwen. Provides up to 18,000 requests per month, and it's just for $10. Now, we've got uh Ollama here, which is probably you know, competing up against Ollama. Ollama, they start at $20, and they're very vague about their actual usage. Um they just they actually don't say anything for the pro, and they say five times more usage than pro. But, you do get that flexibility, cuz I think a lot of people are looking for that flexibility. Now, I was looking at the um T's and C's on this, and I misread it. Basically, I thought I'd read that you can't use it on anything or any automations, or anything that you're not in front of directly using it, you know? Uh and that ruled out OpenClaw. But, they actually explicitly mention that it can be used for OpenClaw. They say Claude code and OpenClaw. It it is all all gravy, baby. The thing about this, if I can find my phone, is that they've managed to get it to 6 gig. There's a lovely app, actually. Um I used to I used to use a local LLM. Have you Do you ever use any sort of local LLMs, or anything like that? No. I I kind of have this tendency to go for the absolute strongest, best ultra max pro ones. Yes, yes. The old the old Maximilian. So, this is a nice app, Locally AI. Locally AI. Offline and private personal assistants. So, you can Is that in focus? You can use just Apple foundation models, but you can also use Qwen 3.5 on on your phone. And this is all completely 100% um local. So, yeah. If you ever want to play with it, I I don't know what phone you have, but it you know, it works fine on this, but it's um It's a good one. It's Yeah, I just set something off like. Nothing's wrong with it. Just wrote a little email there for me. To like no time at all. Um but yeah, anyway. That's actually cool especially for a Chinese model. Yeah, GDPR compliance issue. GDPR, you know. I don't know if you Yeah, I don't know what people want to do on their phones or whatever, but I don't know. Like what What can we ask it? Cuz it's like again having like the world's information in your pocket. Like Let's write a new conversation. Let's say like I don't know. Um Shh uh distance from [snorts] uh uh Earth to Jupiter. I mean, this is more of a Google search to be fair. Yeah. it tells me 484 million kilometers. Took Took 2 seconds. Thing is I don't know. Like Yeah. Like where where's It's like trust me, bro. [laughter] Yeah, uh we we know for a fact that uh like these models if they don't have access to the internet and if they are not thinking uh they perform quite bad. Mhm. Like Mhm. Yeah. It They They are just much better with these two skills. They are kind of like very basic skills that you need or like tools uh that they need to have. Well, it's 4 billion parameters and it is 3 gig. You know, and has vision as well. So, it has vision baked in. Okay. Um Yeah, so I don't know. It's really useful to have these type of models if you um if you want to go back in time and bring your phone with you to restart civilization. Mhm. It really is. [laughter] This is part of a larger conversation on my show, Command AI, which we stream live every single week. We discuss the news and all things related to AI in the world of design and web. Catch us next week and join in the banter. Command AI.