Kimi, GLM and MiniMax, the Chinese alternatives I use

My main tool is Claude, and I also work with OpenAI. But in my day-to-day there are four other names that many people barely know or don't like the sound of: Kimi, GLM, MiniMax and DeepSeek. Here's what I use them for, how they stack up against the usual suspects and, above all, what data I give them and what I don't.

The real reason is less heroic than it sounds: the per-use API price and curiosity in equal measure. I kept reading that there were Chinese models that could code and work as agents almost at the level of the big ones, and I wanted to check that with my own tasks.

I started in early 2025, with DeepSeek, like almost everyone, and from there the others followed: Kimi in mid-2025, GLM shortly after and MiniMax last, in 2026.

Two extremes circulate about these models. One says "they're just as good and much cheaper, anyone paying the Americans is a fool". The other, "don't even think of touching them, they'll steal your data". Neither has been any use to me.

A warning before going on: this isn't a technical analysis or a model ranking. It's my experience, with my tasks, and with all the biases that implies.

Kimi is an AI assistant, similar to ChatGPT, made by the Chinese company Moonshot AI, which also gives that name to its family of models. It's known for handling very long texts and for models, like the K2 series and its successors, designed to work as agents and to code.

I use it in two ways. In the chat, to read and summarise very long documents, such as tender specifications, manuals or specs running to hundreds of pages. And via API, inside an agent harness, for coding tasks with no sensitive data. A harness is the program that wraps the model and gives it tools to act.

What it does well, in my experience: it handles huge documents without losing the thread and, as an agent, it's surprisingly consistent on long multi-step tasks. What it doesn't: in Spanish it writes correctly but flatly, with sentences that sound translated, and sometimes it insists on following a plan even when you tell it the situation has changed.

The example that convinced me most: I gave it a technical manual of about 300 pages and asked about three details hidden in different appendices. It found all of them and cited the section.

First, the naming mess, which took me a while. GLM is a family of language models made by the Chinese company Z.ai, formerly called Zhipu AI. The name comes from General Language Model, and it started out as ChatGLM. So if you read Zhipu, ChatGLM, GLM or Z.ai, you're looking at the same house.

Many of its versions are released with open weights, meaning you can download them and run them yourself.

I use it mainly via API, as a coding model inside agents, and now and then in its chat for quick questions. What for: server maintenance scripts, small internal tools and reviewing my own code.

Have I tried it on my machines? Yes, a small version of GLM on the server with the most graphics memory. It worked, but for what I needed it fell short of the big version via API, which wouldn't fit at home by a long way.

What it does well: coding; on everyday coding tasks it has given me results very close to those of the big paid models. What it doesn't: when the task mixes code with long explanations in Spanish, it feels less polished, and once or twice it has invented a configuration option that didn't exist.

MiniMax is a Chinese company based in Shanghai that many people know only for the videos from Hailuo, its video generator. But behind it there are also text models, the M series, designed for conversation, coding and agent work, as well as voice, music and image models.

What I use from MiniMax is the text model via API, as an alternative for agent tasks, and I've done the odd test with Hailuo for short demo videos.

What surprised me: the speed. On agent tasks with lots of short steps, it responded faster than almost all the others, and in an agent that really shows. What disappointed me: on tasks that need a bit more thought, like reorganising a project with several files, it made hasty decisions that then had to be undone. And Hailuo's videos, very eye-catching, were of no use for anything professional.

DeepSeek is the best known of them all: a Chinese AI company, with its chat assistant, its API and some of its models released openly.

I use it in the chat for one-off questions, via API for reasoning tasks and, locally, with small versions of its open models. On top of that, I've spent ten days testing DeepSeek Harness, its open-source tool for working with agents, and I write about it in Ten days working with DeepSeek Harness.

I treat it separately because it's not just one more model on my list: with Harness it has become a different way of working, and it deserves its own article.

This is what I was really interested in. I did the same tasks with several of them and noted who did best, who was fastest and what it cost me. I'm not quoting published prices; what you see is what I spent.

Writing in Spanish. A text of about 800 words for a client's website, from some notes. Claude was by far the best, with a text I barely had to touch. OpenAI, fine. The Chinese ones, all behind: grammatically sound, but with Spanish that sounds translated.

Summarising a long document. A technical tender specification of about 200 pages. Kimi and Claude tied on quality; Kimi was somewhat faster. GLM and MiniMax missed an important detail or two.

Coding a script. A backup script with email alerts. Claude and GLM got it first time; Kimi, second time; MiniMax needed three rounds. The Chinese ones cost a fraction of what Claude did.

Working as an agent. A task of about twenty steps, migrating a small website from one server to another. Claude finished it unaided. Kimi did too, more slowly. MiniMax was the fastest, but I had to correct two of its decisions.

The honest caveat: these are my tasks, not a general ranking. And models change every few weeks, so this table is out of date before you read it.

What you type into their chat or send through their API travels to their servers and is governed by their terms and by the law of their country. That's the same with any provider, American ones included. But it's worth thinking about, and I've had to think about it.

My actual rule. I give them public texts, tests, my own code with no sensitive data or keys, technical documentation that's already published, and learning tasks. I don't give them client data, internal company information, passwords or anything I'd be embarrassed to see published.

Have I changed my mind? Yes. At first I was more trusting and passed one of them a server log file without checking it. Later I saw it included internal addresses and usernames. Nothing happened, as far as I know, but since then I check everything that goes out to any provider, Chinese or not.

The question of whether it's safe to give data to Z.ai, or any of them, doesn't have a technical answer: you have to weigh it up yourself, with their terms in front of you.

There's an alternative when the model is open: download it and run it on your own machine, and then nothing leaves the house. The limitation is clear: the big versions don't fit on a home computer, and the small ones don't perform the same.

Today I still use them for coding and agent tasks with no sensitive data, where the savings are real and the quality is good enough, and Kimi for reading very long documents. I've gone back to Claude, or OpenAI, for anything written in Spanish for other people, anything that touches client data and agent tasks where a mistake is costly.

Who would I recommend them to? Anyone who wants to experiment and save money on tasks with no sensitive data. Who wouldn't I? Anyone who's going to feed in client data without reading the terms first.

My practical advice, if you're tempted: start with texts you wouldn't mind seeing published. If they work for you there, you'll have time to decide how far to go.