We started Starwisp because good AI shouldn't mean sky-high prices, locked-away models, and your data up for sale. Everything we build is private, ad-free, and priced fairly.
Switch between models from five labs, with no extra apps, logins, or bills.
The principles behind everything we build, starting with Starnexus.
Automatic mode reads each question — a quick reply, code, hard reasoning, a long document — and sends it to the lab that handles that kind of work best.
Every answer follows an ordered chain that spans different labs. If a provider fails, the next one answers instead.
MiMo, Qwen, Gemini, Kimi, and GLM. One login, one interface.
No clutter, no learning curve. Pick a model, ask a question, get an answer.
Your conversations belong to you. We don’t read them, sell them, or use them to train models.
No banners, no tracking, and no sponsored answers.
One subscription for all of them, at a fraction of what five separate AI plans cost. The free tier never expires.
See all plansPower users end up paying for several AI tools at once. Here's what that costs, next to a single Starnexus plan.
Starnexus is built by Starwisp, a small team that didn't think good AI should be locked behind five separate subscriptions. So we put every model in one place, kept conversations private, and made it affordable.
Everything you might be wondering about Starnexus.
Five labs so far: MiMo (Xiaomi), GLM (Zhipu AI), Gemini (Google), Kimi (Moonshot AI), and Qwen (Alibaba). Which models you get depends on your plan. Free includes 3, and Pro includes all 14. New versions get added to Starnexus without you needing another account.
The free plan doesn't expire and doesn't ask for a credit card. It's meant for everyday use: asking questions, drafting things, exploring ideas. Paid tiers exist if you want heavier limits, image generation, or the top reasoning models.
Plans are month-to-month. Move up a tier when you need more, move down when you don't, or cancel from your account settings. You keep access through whatever billing period you've already paid for.
We don't sell what you type, show ads against it, or use it to train models. We don't have models of our own. Your messages do go to the lab that runs the model you picked, and that lab's own terms govern what happens to them there. You can delete any conversation whenever you want.
AI answers should come from what's actually right, not from who paid for placement. Ads push sponsored recommendations and quietly distort what the model says, so we keep them out.
Today there's a web app that works in any modern browser. Native apps for macOS, iPhone, and iPad arrive in September, and Android follows in October, working on most recent phones.
More models, longer conversations, higher limits
Every model, one nexus.
Starnexus is one chat app that puts five leading AI labs behind a single login: MiMo, Kimi, Gemini, Qwen, and GLM. Switch models mid-conversation, keep everything private, and pay one subscription instead of five. There's a free plan, and no credit card to start.
More Starwisp products are in the works.
Starnexus gives you direct access to the leading models from Xiaomi, Moonshot, Qwen, Zhipu AI, and Google. Pick the right one for each task.
Efficient open models focused on reasoning and agents, with a small footprint, fast speed, and strong coding.
Ultra-fast reasoning, coding, and agents.
Efficient 15B-active Mixture-of-Experts built for high-speed reasoning and agentic workflows. Strong on math and software engineering benchmarks while staying cheap and fast.
Omni-modal agent with 1M context at half the cost.
Xiaomi's native multimodal model. It understands text, images, audio, and video, and delivers Pro-level agentic performance at roughly half the inference cost, with a 1M-token context for long documents and multi-step tasks.
Xiaomi's flagship for long-horizon autonomous work.
Xiaomi's most capable model, built for complex software engineering and tasks that take humans days. It has a 1M-token context. Xiaomi reports it sustaining 1,000+ tool calls in a single run, and leading ClawEval, GDPVal, and SWE-bench Pro.
Specialises in long-context reasoning, agentic workflows, and deep tool-use.
The coding model, served faster.
K2.7 Code on a faster serving tier, for rapid iteration in long coding sessions. Included on the Pro plan.
Moonshot's flagship, for long-horizon coding and agent swarms.
Moonshot's most capable model, built for long-horizon coding, agentic workflows with coordinated sub-agents, and deep tool-use over very long contexts. Included on the Pro plan.
Qwen builds one of the world's most-used model families, with strong multilingual, coding, and agentic performance.
The balanced middle of the Qwen line.
A cost-effective step up with stronger reasoning and longer context, for research and heavier daily work. Included on the Plus plan.
The most capable Qwen model.
A sparse mixture-of-experts model that reads images as well as text, with a 1M-token context for long documents and multi-step work. Alibaba puts it at 2.4 trillion parameters. Included on the Pro plan.
Builds the GLM family of general-purpose models, with a dedicated vision line.
Reads images, screens, and documents.
A vision model built to understand images, screenshots, and scanned documents alongside text. Included on the Pro plan.
The Gemini family, engineered for multimodal reasoning at scale across text, images, video, audio, and code.
A fast, cheap default.
Google's fastest model, tuned for instant answers rather than deep thought. Handles everyday tasks and complex agentic workflows with multimodal input. A strong default for most chats.
Google's deepest reasoning, built for hard problems.
Extended thinking mode over the same multimodal core as the rest of the Gemini family: text, images, video, audio, and code, with a 1M-token context. Optimized for software engineering, agentic workflows, and multi-step reasoning.
We started Starwisp, a California-based startup. Every AI company charges its own subscription, and if you want to use more than one model you end up paying two, three, sometimes four separate bills. On top of that, switching between apps just to try a different model is tedious: a question for Gemini here, a prompt for Kimi there, another tab for Qwen. It shouldn't be this complicated or this expensive.
So we built one app that brings MiMo, Kimi, Gemini, Qwen, and GLM together in one place. Switch between them instantly, no extra tabs, no extra logins. One subscription covers everything, and there's a free plan too, with no credit card and no trial that expires. We also believe ads have no place in an AI product. Your conversations are yours. We don't mine your data, sell it, or use it to train models. Privacy is how Starnexus works from day one, not something we bolt on later.
We built this because we wanted it ourselves.
Starwisp is a real design. In 1985, physicist Robert L. Forward proposed it: an interstellar probe made of an ultra-thin wire mesh, light enough that a microwave beam could push it to a fifth of the speed of light, with no rocket needed. Tiny circuits were built directly into the wires so the whole thing could still think and send data home.
Same idea, but we're the other half of it. The real Starwisp only reached the stars because a massive transmitter on Earth was doing all the heavy lifting. That's us. We run the models and the infrastructure behind everything. The app you hold is the light part, built to reach all of it.
Icons come from the Solar Icon Set, used under the Creative Commons Attribution 4.0 license. We use the line-duotone variant, modified for this site: resized, recolored to match our palette, and with the secondary tone set to 0.45 opacity.
Starnexus connects models from Xiaomi (MiMo), Zhipu AI (GLM), Google (Gemini), Moonshot AI (Kimi), and Alibaba (Qwen). All model names and trademarks belong to their respective owners.
Announcements and notes from the team.