Z.ai’s GLM-5.3-Flash is cheap, good, and served on Chinese chips
The New Stack Frederic Lardinois ● Covered by 2 sources
Z.ai just revealed GLM-5.3-Flash, a cheap open model that was hiding as ox-alpha. It’s fast enough to matter, and Z.ai says it’s being served on Chinese chips.
Based on reporting by The New Stack, Frederic Lardinois — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
For the last few days, ox-alpha was the mystery model getting attention on OpenRouter. Now the mask is off: it’s Z.ai’s GLM-5.3-Flash, a 320 billion-parameter hybrid model with 18 billion active parameters, built specifically for ultra-low-cost inference.
Z.ai made the weights public on Wednesday and put them on Hugging Face under the MIT license. The model is already live on several inference platforms, including OpenRouter, where it’s listed at $0.075 per million input tokens and $0.25 per million output tokens, with those prices reflecting a 50% discount.
The benchmark story is a little less breathless than the hype suggested. GLM-5.3-Flash is not at Claude Fable 5 level, despite some early chatter, but it can mostly keep pace with Claude Opus 4.8 and OpenAI’s GPT-5.6 Terra when it’s switched to max-effort reasoning. On the Artificial Analysis Intelligence Index, it lands at 57 points, alongside GPT-5.6 Terra, Gemini 3.7 Flash, Muse Spark 1.2, and Qwen 3.8 2.4T A95B.
Where it gets more interesting is agent work. Z.ai says the model does especially well when driving AI agents, which may be a better proxy for real use than leaderboard theater. It also handles multimodal inputs like images, videos, and files, and Z.ai says it trained it to improve presentation building, website building, document work, spreadsheets, and dashboards.
There is a tradeoff. The model is chatty and burns through tokens, which would be annoying if the serving cost were ordinary. But Z.ai’s whole pitch is that the serving cost isn’t ordinary. The company says it was able to serve 100 trillion free tokens per day, according to OpenCode, and that it did so on Chinese AI chips. Z.ai says it got a 3× boost over its initial baseline on the same hardware, with per-token cost and hardware efficiency comparable to mainstream NVIDIA GPUs.
Under the hood, Z.ai says it built its own inference engine on top of SGLang, helped along by an infrastructure agent powered by its flagship GLM-5.3 model. The model itself was trained on a 30 trillion multimodal pre-training corpus and uses a hybrid mix of linear and sparse attention. Z.ai says that cuts compute by 3x and KV cache size by 4.4x versus the full GLM 5.3 model, which matters a lot when the context window runs to 1 million tokens.
My take — AI-written commentary, not fact-checked reporting
This is the kind of release that should worry the U.S. labs more than any leaderboard screenshot. Cheap open weights plus real serving scale beats prestige every time, and China’s chip story is no longer just a footnote. The market loves a miracle model; it usually pays more attention when the invoice is tiny.
Read more about this at: The New Stack
