VibePager · Models
The models are your machine’s, not ours
Anthropic, OpenAI, Google DeepMind and xAI on one side; DeepSeek, Moonshot AI, Z.ai, Alibaba and MiniMax on the other. VibePager hosts none of them — it reads the model list out of the agent running on your own computer, which is the only way a list stays true in a field that turns over every few weeks.
Updated September 2026
Why the list is not in the app
Any model list written into a phone app is wrong within a month. Between them, the open-weight labs shipped several generations in 2026 alone, and the closed labs are no slower. A hardcoded list fails in the worst way: it offers you a model your machine cannot actually run, and the conversation fails somewhere you cannot see.
So the picker asks the agent. Whatever OpenCode, Kilo Code, Kimi, Gemini CLI, Codex or Claude Code offers on your machine, with your keys, is what shows up — with search, because some of those lists are long.
The families, and what they are called right now
Read the version column carefully: it is filled in differently for the two halves of the table, on purpose. The open-weight labs publish versioned checkpoints you can name, pin and download, so naming them is useful. The closed labs rename and re-tier constantly — while this page was being written, three reference sites disagreed about which OpenAI and Google flagship was current that week. So the closed families get tiers, not numbers.
| Family | Who builds it | Weights | Named versions (Sept 2026) | Usually reached through |
|---|---|---|---|---|
| DeepSeek | DeepSeek AI — Hangzhou, China | Open | DeepSeek V4, in Pro and Flash variants | Their API, a gateway, or self-hosting |
| Kimi | Moonshot AI — Beijing, China | Open | Kimi K3, and K2.5 before it | Kimi CLI, self-hosting, or a gateway |
| GLM | Z.ai (Zhipu AI) — Beijing, China | Open | GLM-5.3 and GLM-5.2 | The Z.ai GLM coding plan, a gateway, or self-hosting |
| Qwen | Alibaba Cloud — Hangzhou, China | Open | Qwen 3.8 Max and Qwen 3.6 | Qwen Code, a gateway, or self-hosting |
| MiniMax | MiniMax — Shanghai, China | Open | MiniMax M3 | Their API or a gateway |
| Claude | Anthropic — San Francisco, USA | Closed | Opus / Sonnet / Haiku tiers | Claude Code, or an Anthropic key in another agent |
| GPT | OpenAI — San Francisco, USA | Closed | flagship / mini tiers | Codex CLI, or an OpenAI key |
| Gemini | Google DeepMind — Mountain View, USA | Closed | Pro / Flash tiers | Gemini CLI, or a Google key |
| Grok | xAI — Palo Alto, USA | Closed | Grok and Grok Code tiers | An xAI key in an agent that accepts one |
| Mistral | Mistral AI — Paris, France | Open and closed | Codestral and the Large tiers | An API key in a model-agnostic agent |
The labs, briefly
Worth knowing who is who, because the split is not subtle and it decides what you can actually do with a model.
- Anthropic (Claude), OpenAI (GPT, Codex), Google DeepMind (Gemini) and xAI (Grok) keep their weights closed. You rent access through an API. They set the price and the rate limits, and the model can be retired from under you.
- DeepSeek, Moonshot AI (Kimi), Z.ai — formerly Zhipu AI — (GLM), Alibaba Cloud (Qwen) and MiniMax publish open weights. You can use their hosted API, a gateway, or download the model and run it yourself. That last option is the one that cannot be taken away.
- Mistral AI sits between the two, publishing some weights openly and keeping others behind an API.
The Chinese labs are the reason this page has a version column at all. Between them, DeepSeek, Moonshot, Z.ai, Alibaba and MiniMax have shipped the bulk of the serious open-weight coding models — at prices that make it reasonable to leave an agent running on a long job, which is exactly the case a phone is for.
Running a specific model through a specific agent
This is the question most people are really asking — not "which model is best" but "can I run this model in that agent". The answer is almost always yes, and it has nothing to do with VibePager: it is the agent’s provider configuration.
- Model-agnostic agents — OpenCode and Kilo Code — are built for exactly this. Point them at an endpoint and the model shows up in their list, whether it is DeepSeek, MiniMax, GLM, Qwen or something you host yourself.
- Gateways such as OpenCode Zen, OpenRouter or a self-hosted router put many models behind one key. The agent sees one provider; you see the whole catalogue.
- First-party agents — Claude Code, Codex, Gemini CLI — are tied to their own lab’s models by design, though some accept a custom base URL.
- Self-hosted endpoints work like any other provider. If the agent on your machine can reach it, the phone lists it.
Whatever you end up with, the phone shows that list — read from your machine, with search. VibePager never decides which models you may use, and never adds one you cannot reach.
A dated snapshot — and the point of dating it
Everything in the version column above is from September 2026. Six months earlier the same column would have said Kimi K2, GLM-4.6, DeepSeek V3 and Qwen3 Coder — different names, same products, same companies. Nothing about the tools changed; only the numbers did.
That is measurable, not a hunch. Searching for the old names and the new ones side by side in September 2026, the new ones had already overtaken: glm 5.2 free was ahead of glm 4.7 free, and the strongest model queries of all were for versions that did not exist half a year before — minimax m3 price, deepseek v4 pro pricing, glm 5.3 pricing.
The same goes for benchmark rankings and prices. Whichever model tops a coding leaderboard this month will not next month, and this page deliberately quotes no prices: they move faster than it can be republished, and a stale price is worse than no price. Check the provider.
Cheap model, long job, walked away
There is a pattern worth naming, because it is where a phone earns its place. Open-weight models cost little enough that you will happily set one on a long, boring refactor — the kind of job you start and leave. Which means the agent will stop to ask permission at minute nine, while you are somewhere else, and sit there until you come back.
The cheaper the model, the more likely you are not watching. That is the case the notification exists for.
What VibePager does not do
- It does not host or serve models. There is no inference here and no token markup.
- It does not need your API keys. They stay where they already are — in the agent’s own configuration on your machine.
- It does not route your prompts through a model. The relay forwards encrypted bytes between your phone and your computer and cannot read them.
Questions
- Which models can I use from the phone?
- Whichever ones the agent on your machine offers. If OpenCode on your laptop can reach a model, the phone lists it. Nothing is hardcoded into the app.
- Can I run open-weight models like Kimi, GLM, DeepSeek, Qwen or MiniMax?
- Yes, through an agent that can reach them — OpenCode and Kilo Code are model-agnostic, and Kimi CLI runs the Kimi family directly. Self-hosted endpoints work the same way: if the agent can reach it, so can the phone.
- Do I give VibePager my API keys?
- No. Keys live in the agent’s configuration on your own machine, exactly where they are now.
- Is there a markup on tokens?
- No. VibePager is a flat subscription for the remote control; you pay your model provider directly, as you do today.
- Which model is best for coding?
- It changes every few weeks, which is the honest answer and also the reason the picker reads from your machine. The practical split most people land on: a cheap open-weight model for long or repetitive work, a premium closed model for the hard parts.
VibePager runs the agents on your own machine and puts the controls on your phone.