GLM-5.3-Flash, tested as Ox Alpha, released with hybrid architecture & scaling efficiency
Z.ai has introduced GLM-5.3-Flash, the first natively multimodal model in its GLM-5 series. The model has 320 billion total parameters, with 18 billion active parameters. Z.ai says it outperforms GLM-5.2 across benchmarks and real-world workloads at one-tenth the price. The company also says it approaches Claude Opus 4.8 on coding and agentic benchmarks.
Building on those performance claims, GLM-5.3-Flash combines sparse and linear attention in a hybrid architecture. Z.ai says the design lowers long-context serving costs while retaining precise long-context capabilities. The model also uses Manifold-Constrained Hyper-Connections, or mHC, to improve scaling efficiency. Its pre-training used a 30-trillion-token multimodal corpus.
For million-token contexts, Z.ai introduced IndexPool to reduce indexer latency and memory overhead. The technique compresses four indexer key vectors into one through weighted pooling.
Before the release, Z.ai tested the model anonymously as Ox Alpha on OpenCode and OpenRouter to collect user feedback, and it quickly became the most popular model of the week. The model weights of GLM-5.3-Flash are publicly available on HuggingFace.
- AI Chatbot
- Free
- Open Source
External links
- GLM-5.3-Flash: Frontier Intelligence, Flash CostZ.ai • Official source
- The Powerful Stealth AI Model ‘Ox Alpha’ Is Now GLM-5.3-Flash, and You Can Use ItCNET
- The cow comes home: China’s Internet celebrates Z.ai claiming ownership of Ox AlphaBusiness Insider
- Sanctioned Chinese Chips Just Served 62 Trillion AI Tokens at Frontier ScaleTech Times
- LM Studio Bionic adds GLM-5.3-Flash support9to5Mac
- GLM-5.3 is here with advanced cyber capabilities — and reportedly already found a 'serious vulnerability' in CursorVentureBeat
Comments
To be fair, a model become popular if its relevant and cheap, since it'll be used for many tasks and will burn a lot of token. Larger and more powerful models will burn much less tokens because they are used for a smaller set of tasks (except if you work at OpenAI or Anthropic and you're allowed to burn as much as possible to fake demand) even if theses tasks are harder (they could burn less than 10% of the total tokens).
Because agents don't choose by name/brand but prefers to use some model every time this model is doing better than the others. And this Ox model has been voluntary much cheaper (and in sale) than the others because of the user content collected to post-train it. But it has only burn 20% of the total tokens on OpenRouter, pretty much like DeepSeek V4 Flash a month ago.
Still, Ox/GLM is a strong contender for small and medium tasks (as a "Flash" model, for what is means), has solid throughput, decent context size and has a great cache hit. Still needs a decent rack of GPUs to run though.