GLM-5.3-Flash: MIT open weights that read your documents

Most of the boring work in a small business is not text. It is a scanned invoice, a supplier PDF, a screenshot of a portal nobody has an API for. Z.ai released GLM-5.3-Flash on August 26 under an MIT license — the first natively multimodal model in the GLM-5 line, and the first time cheap open-weight multimodal inference lands in the range where you would actually point it at a stack of documents.

What actually happened

Per Z.ai's model docs, GLM-5.3-Flash is a mixture-of-experts model with 320B total parameters and 18B activated per token, a 1M-token context window, and native visual input covering images, video, and files. Thinking mode is always on — it cannot be disabled.

The engineering detail that matters for anyone self-hosting is the attention rework. Against its predecessor, Z.ai reports attention computation cut by 3.01x and KV cache size by 4.44x. KV cache is what eats your GPU memory on long agent runs, so a 4x reduction changes how many concurrent sessions fit on the same box.

On price, Artificial Analysis lists the hosted API at $0.15 per million input tokens and $0.50 per million output, scores it 57 on its Intelligence Index against a median of 28 for comparable models, and measures output at 50.2 tokens/second — slower than the 67 t/s median. Weights went to Hugging Face under MIT the same day.

Z.ai's own Code Bench puts it at 29.0 against Claude Opus 4.8's 29.5 at maximum reasoning effort. That is a vendor's benchmark on a vendor's harness. Treat it as a reason to test, not a result.

Why cheap multimodal matters for your business

The economics of document work have been stuck. Vision-capable frontier models are priced for occasional use, so most operators run OCR, get 85% accuracy on anything with a table in it, and put a human on the rest. At $0.15 per million input tokens with weights you can host, the arithmetic flips: you can afford to send the whole page, twice, and cross-check.

Pick one workflow and measure it. Supplier invoices into line items. Packing slips into receiving records. Screenshots of a vendor portal into a nightly sync. Run fifty real documents — your ugly ones, not clean samples — through your current process and through GLM-5.3-Flash, and count the fields a human had to fix. That number is the whole decision.

The 50 t/s output speed is the honest catch. This is not the model for anything a customer is waiting on. It is the model for the queue that runs at 2am, where being slow costs nothing and being cheap compounds. And MIT means the version you validate is the version you keep — no deprecation email, no retired alias, no repricing in November.

Sources: Z.ai docs, Artificial Analysis.