If you were poking around OpenRouter or OpenCode last week and ran into a model called ox-alpha
, you probably noticed the same thing everyone else did: it was fast, it held up on long agentic runs, and it kept finishing tasks that models in its apparent weight class had no business finishing. It became the most-used model of the week before anyone knew who built it.
It’s Z.ai. The model is GLM-5.3-Flash, and it’s available in Kilo right now across the IDE extension, the CLI, and Cloud Agents.
What it actually is #
GLM-5.3-Flash is 320B total parameters with only 18B active, and it’s the first natively multimodal model in the GLM-5 line. Z.ai prices it at roughly a tenth of GLM-5.2 while beating it more or less everywhere, which is the part that made us pay attention.
The architecture is where the savings come from. Z.ai went with a hybrid of linear and sparse attention: linear attention handles local dependencies through state modeling, sparse attention pulls in global context through a lightweight indexer. On top of that they added IndexPool, which compresses four indexer key vectors into one so the indexer doesn’t blow up memory at 1M-token context. Against GLM-5.3, that combination cuts attention compute by 3x and KV cache size by 4.4x.
Compared to the GLM-4.5 generation, the numbers are almost funny. Similar total parameter count (320B vs 355B), but roughly half the active parameters and half the layers. Fewer moving parts per token, more capability out the other end.
The benchmarks #
Z.ai’s own numbers put GLM-5.3-Flash near Claude Opus 4.8 on coding and agentic work:
Benchmark GLM-5.3-Flash GLM-5.2 Opus 4.8 Terminal Bench 2.1 84.3 81.0 85.0 DeepSWE v1.1 63.4 46.2 58.0 Toolathlon Verified 78.4 59.9 76.2 AutomationBench v1.0.6 48.8 26.2 41.0 GDPval-AA v2 1773 1504 1582
The jump from GLM-5.2 on DeepSWE (63.4 vs 46.2) and AutomationBench (48.8 vs 26.2) is not a rounding-error improvement, it’s a different tier of model. On Z.ai’s internal Code Bench, run through Claude Code, max-effort GLM-5.3-Flash lands at 29.0 against Opus 4.8’s 29.5.
And on the Artificial Analysis Intelligence Index v4.1.1, it scores 57 at about $0.045 per task at the discounted rate. That number is the whole story. A score in that range used to cost roughly ten times as much.
Benchmarks are benchmarks, though. The more convincing evidence is that thousands of developers used it blind for a week, had no idea who made it, and kept coming back.
Vision that’s actually useful for coding #
The multimodal part isn’t a checkbox feature. Z.ai trained it specifically around visual self-verification, meaning the model renders its own frontend output, looks at it, notices the layout is broken, and fixes it without you pasting a screenshot and saying “this looks wrong.”
That matters for the work where correctness isn’t just “does it compile.” Frontend, game dev, 3D, anything with a canvas. A component can pass every test and still render as a pile of overlapping divs. A model that can see the result closes a loop that used to require a human in the middle.
It also carries over to non-code work: documents, spreadsheets, dashboards, slides. GLM-5.3-Flash posts 62.4 on OfficeQA Pro and 78.0 on Chartography with tools, both ahead of Opus 4.8.
One more thing worth flagging #
Z.ai served this entire launch on a large-scale cluster of Chinese AI chips. They built a dedicated inference engine on top of SGLang, leaned on W8A8 quantization and a disaggregated Encode-Prefill-Decode architecture, and got a 3x end-to-end improvement over their own baseline on the same hardware. Their claim is that per-token cost and hardware efficiency now sit in the same range as mainstream NVIDIA GPUs.
They also mention that a GLM-5.3-powered infrastructure agent helped their engineers write and optimize the kernels for it. The model helped build the stack that serves the model. Make of that what you will, but it’s a real data point on where agentic work is landing.
Where it fits in your Kilo workflow #
We’ve been running it internally for a few days. Some early impressions:
It’s a great default. At this price and this speed, GLM-5.3-Flash is the model you leave selected for most of the day. Code mode, routine implementation, refactors, test scaffolding. You stop thinking about the meter.
It’s very good under parallel agents. Cheap tokens plus fast generation means you can fan out five or six agents on a feature without the cost math getting uncomfortable. Long-running Cloud Agents are the same story: the 1M context and the low serving cost mean a session can grind for a long time before anything hurts.
Code Reviewer is an obvious fit. Reviews are high-volume, bounded, and repetitive, which is exactly what a fast model with strong tool use is for. Point Kilo’s Code Reviewer at GLM-5.3-Flash and you get PR feedback on every open PR without the per-review cost creeping up.
Frontend work is where the vision pays off. Give it a design reference or let it check its own render, and you spend a lot less time describing what’s broken.
Try it #
GLM-5.3-Flash is in the model picker now, everywhere Kilo runs. Open the extension in VS Code, JetBrains, or the CLI, and select it. If you’d rather run it yourself, Z.ai published the weights on HuggingFace with SGLang, vLLM, and TokenSpeed support.
Tell us how it goes in Discord. We’re especially curious whether the vision loop holds up on real frontend work, since that’s the piece benchmarks are worst at capturing.