On October 16, OpenCode posted that GLM 5.3 Flash was now free, and I read the news wrong. I thought Z.ai had given away its top model. Same family, same number, so how different could it be? Quite different, it turns out. The free one is Flash, the cheaper model, and the full GLM 5.3 still costs about $1.40 per million input tokens and $4.40 per million output tokens through OpenCode Zen.
I caught my mistake from a hands-on writeup by Apidog. They installed a clean copy of OpenCode the same day, used the model with no login and no API key, and had it write and test a Node.js file. It worked. So the free access is real, but it has conditions, and those conditions decide whether this is useful for you or just a good headline. I’ll go through them first, then the question sitting under all of this: can you cancel a Claude or ChatGPT subscription because of it? After that comes the bigger one, which is how close we are to running something this good on our own machine.
What GLM 5.3 Flash actually is
Z.ai, also called Zhipu AI, released it on August 26, 2026. The name misleads people, because “Flash” usually means a small model. This one isn’t small. It’s a mixture-of-experts model with 320 billion parameters in total, but only 18 billion are active for each token. So Z.ai gets a big pool of knowledge and a small running cost, which is the whole trick.
Here are the specs from Z.ai’s model card and the providers that host it:
- 320B total parameters, 18B active per token
- 1,048,576 token context window, with up to 131K tokens of output
- Natively multimodal, so it takes images and video as well as text, and it’s the first model in the GLM-5 series to do that
- A hybrid of sparse and linear attention, a first for GLM, meant to keep long-context accuracy while cutting the serving cost
- Open weights under the MIT license
- Low, High and Max thinking modes, with Max as the default according to Unsloth
On the API it’s $0.15 per million input tokens and $0.50 per million output tokens, with cached input at $0.03. NovitaAI was listing half of that as a promo when I looked. You’ll also see the model called ox-alpha in some places, which was its codename before launch.
What is it good at? Z.ai pitches it for coding agents, the kind that look through a codebase, change several files, run terminal commands, run the tests and go around again. The hosting pages also mention document analysis, research help and multi-step automation with tools. And since it reads images and video, you can hand it a screenshot of a broken page instead of describing the problem in words, which older GLM models couldn’t do.
The free version has limits
The free model in OpenCode has the ID opencode/glm-5.3-flash-free, and it isn’t the same product as the paid API. It gets 200K of context and 32K of output, against 1M and 131K on the paid side. For one task that’s plenty, but you’ll feel it on a long agent session. Apidog also found that it only works inside OpenCode, because when they called it from another client they got a FreeTierError saying the free tier can only be used from within OpenCode. So if you were hoping to plug it into Cline or Claude Code for free, no. OpenCode’s own docs say the free access is for a limited time, and nobody has said when that ends. The history isn’t comforting either, since glm-4.7-free and glm-5-free are both marked deprecated now, and back when this same model was still called Ox Alpha, the free week lasted six days.
I’d treat it as a trial, not a plan.
Setting it up will only take a minute. Open a project, start opencode, type /models and pick GLM-5.3-Flash Free. If you want to check it’s available first, opencode models --refresh and a grep for "free" should list it. Apidog counted eleven free models when they did it.
How good is it, really
Z.ai’s numbers look strong. On Terminal Bench 2.1, which tests an AI agent doing real work in a terminal, GLM 5.3 Flash scores 84.3. The older GLM 5.2 scores 81.0. Claude Opus 4.8 gets 85.0, Gemini 3.7 Flash gets 85.8 and GPT 5.6 Terra gets 87.4. That’s close to the top from a model that’s free this week and cheap after. Against DeepSeek V4 Flash, the numbers passed along by CometAPI show 63.4 on DeepSWE versus 54.4, and 84.3 versus 82.7 on Terminal Bench.
Now the part I don’t want to skip. Those are Z.ai’s own numbers, so it’s a company grading its own exam.Independent trackers are less excited. Toolhalla looked at a pinned benchmark table and said GLM 5.3 Flash does not outrank GLM 5.2 there, while Cloudprice ranks it 9th on its intelligence index and 17th on coding. Both can be true. A model can be great at terminal tasks and only decent at everything else.
And the comparison list itself is a problem. Z.ai measured against Opus 4.8 and GPT 5.6 Terra, but newer ones are already out. CometAPI’s site banner says Claude Opus 5.5 and the GPT-6 models are live. So “rivals Opus” means rivals last generation’s Opus.
I’m oversimplifying a bit, though, because terminal benchmarks test one kind of work. Writing and research are different skills, and I haven’t seen good numbers for them on this model yet.
Can it replace a Claude or GPT subscription?
My answer is 70-30. For coding agents, mostly yes. For the subscription itself, no, not yet.
Take the cost first. A heavy coding day might use 5 million input tokens and 1 million output tokens, which is my rough guess and not a measurement. At $0.15 and $0.50 per million, that day costs $1.25. If you coded that hard every single day for a month, it’s about $37. Compare that with what you pay now, and the API route looks very good for people who live in a terminal.
But a subscription isn’t only a model. It’s an app on your phone, voice mode, file uploads, memory, connectors to your email and calendar, and a model that gets better without you touching anything. GLM 5.3 Flash gives you none of that by itself. You’d be gluing a setup together with OpenCode or something like it, and if you enjoy that, great. If you just want to open an app and ask a question, you’ll be annoyed by day three.
Some people say benchmarks are all that matters and the rest is packaging. I don’t buy that, because a benchmark can’t tell you how a model behaves on your messy repo with its three-year-old config files.
I’d keep a paid plan for the days when a wrong answer costs hours, like a nasty bug across a huge codebase. Opus 5.5 and the newest GPT are a generation past what Z.ai tested against, and that gap is exactly where the hard problems live. I haven’t put GLM 5.3 Flash through a real project of mine yet, so this is a judgment from the numbers and not from my own bad afternoon with it.
How close are we to a good local model?
You can run it locally, but only if your machine has a lot of memory.
The smallest GGUF file for it is 86.7 GiB, according to Toolhalla, which added up the file sizes in the Unsloth repo on October 9. That’s file size, not measured memory use, and Toolhalla says plainly that it hasn’t run the model on any of the machines in its post. So treat the number as a floor. A 24GB RTX 3090 can’t hold the weights, and an NVMe drive doesn’t add VRAM, no matter what some forum posts say. Unsloth says a 3-bit version runs on 128GB of RAM,and it later reported its local runs got 1.6 to 3.4 times faster with multi-token prediction. That’s Unsloth’s own claim and I haven’t seen anyone else test it. Support landed in upstream llama.cpp on September 30, and Unsloth Desktop can load it if you don’t want to use a terminal. I don’t know how much quality the 3-bit version loses, and I haven’t found a good test of it yet.
Small tangent. People say “runs on a laptop” and mean a 128GB MacBook Pro, which isn’t the laptop most of us carry.
The 18B active parameters matter here, I think. Speed depends mostly on how much has to be read from memory for each token, so once the weights fit, it should feel usable. The hard part is getting 87 GB or more to fit at all. The wall is memory capacity, not raw compute. Unsloth lists Qwen3.8-Flash at around 75GB of RAM, so that’s the same class of machine.
And the pace is real. Z.ai shipped GLM 5.3 Flash 71 days after GLM 5.2, and Unsloth says it beats 5.2 at one-tenth of the price. Every couple of months a cheaper model catches the last big one. My guess, and it’s only a guess, is that in about two years a 64GB machine will run something at this level well enough for most coding. Memory prices could easily make me look silly.
Right now the subscription cycle isn’t going anywhere. It just has a cheaper competitor every few months.
One loose end if you want to serve it yourself: the Vultr page says you need vLLM 0.27.0 or newer, so check your version before you download 87 GB and find out.
What I’d do this week
Install OpenCode, run /models, pick GLM-5.3-Flash Free and give it one task you’d normally give Claude or ChatGPT. Pick something real, like a bug you already fixed, so you know what the right answer looks like.
Don’t cancel anything yet. If the free tier disappears next week you’ve lost nothing, and you’ll know whether the $0.15 version is worth wiring into your own setup.