GLM-5.3-Flash review
The fast, low-cost tier of Z.ai's GLM-5.3. Strong coding scores and free to download.
Very strong coding scores for a very low price ($0.15 / $0.50 per million tokens). It is free to download under the MIT license, but far too big for a normal computer, so most people will use it online.
What is GLM-5.3-Flash?
GLM-5.3-Flash is the fast tier of Z.ai's GLM-5.3 family, released on 26 August 2026 after a week of anonymous testing on OpenRouter under the name "Ox Alpha". It is built for coding and long agent tasks, accepts text, images and video, and is released under the MIT license.
Key features
- High coding scores: 84.3 on Terminal-Bench 2.1 (Z.ai's test)
- Cheap: $0.15 per 1M input tokens and $0.50 per 1M output
- Free to download under the MIT license
- Reads about 1M tokens and writes up to 131,072
- Takes text, images, video and files
Specs and price
| Made by | Z.ai (Zhipu AI) |
| Released | 26 Aug 2026 |
| Size | Reported as 320B parameters, about 18B at work per word (sources disagree) |
| License | MIT: free, business use allowed |
| API price | $0.15 in / $0.50 out per 1M tokens, repeated input $0.03 |
| How much it can read | About 1M tokens |
| Longest answer | 131,072 tokens |
| Input | Text, images, video and files |
| Computer needed | Full version about 306 GB. Compressed versions need a 128 GB+ machine |
Test scores
| Test | Score | Published by |
|---|---|---|
| Terminal-Bench 2.1Coding agents working in a terminal | 84.3 | Z.ai |
| DeepSWE v1.1Long, real coding tasks | 63.4 (GLM-5.2: 46.2) | Z.ai |
| Humanity's Last ExamVery hard expert questions | 55.3 | Z.ai |
| Artificial Analysis indexOverall score across many tests | 41.8 | Independent |
Scores published by Z.ai are the maker's own numbers. We will add our own results in the hands-on review. Ignore the viral "80% on DeepSWE" figure. It came from a 10-task subset.
How to download and run GLM-5.3-Flash
- Easiest: use it on z.ai, or through the Z.ai API.
- Or use
z-ai/glm-5.3-flashon OpenRouter, where many providers serve it. - To run it yourself, download it from Hugging Face (zai-org). The full version needs a multi-GPU server. Unsloth's 1-bit and 2-bit versions fit a 128 GB machine with Unsloth's llama.cpp branch.
Who it is for
Good fit
- Developers running lots of coding-agent jobs on a budget
- Apps that need long context cheaply
- Companies that want MIT-licensed models they can self-host
Skip it if
- Laptop users who want it offline: it is too big
- Anyone who needs the best overall quality: it scores 41.8 on the independent index against 58 for Claude Opus 5.5
Pros and cons
Pros
- Very low price
- High coding-agent scores from Z.ai
- MIT license
- Reads text, images and video
Cons
- Too big for normal computers
- Independent overall score is much lower than Z.ai's coding numbers suggest
- The launch discount ended on 9 September
- Sources disagree on its exact size
How it compares
Z.ai's own coding scores rival far pricier models, but the independent Artificial Analysis index puts it at 41.8, well behind Claude Opus 5.5 (58) and Gemini 4 Argon (about 53). Per token, it costs a small fraction of either.
Hands-on review In testing
We are running GLM-5.3-Flash on real tasks. Results will be added to this page, at the same address.
How it got here
| Date | What happened | Source type |
|---|---|---|
| Sep 9, 2026 | Half-price launch offer ends.Codersera, GLM-5.3-Flash specs, pricing and MIT weights | Confirmed |
| Aug 26, 2026 | Z.ai reveals it as GLM-5.3-Flash, under the MIT license.OpenRouter, GLM 5.3 Flash model page | Confirmed |
| Aug 20, 2026 | An anonymous model called Ox Alpha appears on OpenRouter.Codersera, GLM-5.3-Flash specs, pricing and MIT weights | Leak |
Frequently asked questions
Is GLM-5.3-Flash free?
The model is free to download under the MIT license. Using Z.ai's API costs $0.15 per million input tokens and $0.50 per million output tokens. The half-price launch offer ended on 9 September 2026.
Can I run GLM-5.3-Flash on my computer?
Only on a very large machine. The full version is about 306 GB, and the smallest compressed versions need about 128 GB of memory.
What was Ox Alpha?
The anonymous name GLM-5.3-Flash used on OpenRouter for a week before Z.ai announced it on 26 August 2026.
Is GLM-5.3-Flash good for coding?
Z.ai reports 84.3 on Terminal-Bench 2.1 and 63.4 on DeepSWE, which is strong for its price. Independent overall scores are lower.
Related models
Flint
RecursiveFlint is out. Launch facts from official sources are below, and our hands-on review follows.
Kolibri
Aleph AlphaRuns cheaply on-premise, which companies with strict data rules need.
Nemotron 3.5 Lightning
NVIDIAOnly 3B of its 30B parameters work at a time, so it is one of the fastest capable models on a single GPU.