Z.ai's GLM-5.3-Flash Gets Near Opus at a Tenth of the Price
A 320-billion-parameter open-weight model with a million-token context window, released under an MIT licence at $0.15 per million input tokens. It spent a week on leaderboards under a fake name first.
Chinese lab Z.ai has released GLM-5.3-Flash, the cheapest capable coding model it has shipped and the first in the GLM-5 family to handle images and video natively rather than through a bolted-on converter. The weights are on Hugging Face under an MIT licence, which is about as permissive as software licences get, and the hosted version costs $0.15 per million input tokens and $0.50 per million output tokens.
The specifications: 320 billion total parameters with 18 billion active per token, and a context window of 1,048,576 tokens. Parameters are the internal numbers a model learns during training; a token is roughly a word fragment; the context window is how much text the model can hold in mind at once, and a million tokens is comfortably a long novel or a very large codebase. The gap between 320 billion stored and 18 billion used is a mixture-of-experts design, where the model keeps many specialised sub-networks and only wakes the handful relevant to each piece of text.
On benchmarks, Z.ai reports 84.3 on Terminal-Bench 2.1 against Claude Opus 4.8’s 85.0 and GPT-5.6 Terra’s 87.4, and 63.4 on DeepSWE v1.1 against GLM-5.2’s 46.2. Independent testing by Artificial Analysis scores it 57 on their Intelligence Index, well above the median of 27 for open-weight models of comparable size, though it is slow at 48.7 output tokens per second and noticeably verbose. Vision is the weak spot: it trails Gemini 3.7 Flash on image benchmarks.
What’s actually going on here: the model spent its first week running anonymously as “Ox Alpha” on OpenCode and OpenRouter, which is a now-common way of getting honest reactions before the branding kicks in. More interesting is where it ran. Z.ai says the entire preview was served on domestically produced Chinese AI chips, with a custom serving engine that reports a threefold end-to-end improvement. That is the part worth watching. The export-control story has always assumed Chinese labs are bottlenecked on Nvidia hardware, and this is a lab claiming it shipped a near-frontier model without it. Treat the claim as reported rather than verified, and remember all the headline benchmark numbers are Z.ai’s own with per-test setups that differ.
What this means for you: if you use AI through a chat window, the practical effect is downward pressure on prices everywhere, which reaches you eventually. If you pay for API access, this is genuinely cheap for the quality: roughly $0.045 per task on the discounted tier. If you were hoping to run it at home, temper that. The default FP8 checkpoint is around 306 GiB before the memory the model needs while running, and the current serving path wants Nvidia Hopper chips or newer, so this is an eight-GPU-node model rather than a desktop one. For most people the API is the realistic door, and the open licence mostly matters because it means nobody can take the model away from you later.
Sources
Source: https://z.ai/blog/glm-5.3-flash
Google Built a Tiny AI Model That Reads Glucose Traces
GlucoFM has 720,000 parameters, roughly a millionth the size of a chat model, and beats larger rivals at spotting metabolic patterns in continuous glucose monitor data.
