Z.ai's Flash Model Is 320B Parameters, At One-Tenth Its Own Price

Z.ai published weights for GLM-5.3-Flash on Hugging Face on 26 August, under an MIT licence. The model had been circulating anonymously as a stealth endpoint before the reveal. Three details in the model card change how the release should be read.

"Flash" is 320 billion parameters

The card gives 320B total parameters with 18B active per token. In every other vendor's naming convention, Flash, Mini, Turbo and Lite denote a small model. Here it denotes a sparse one: a mixture-of-experts design where only 18B of the 320B are engaged for any given token. The inference cost profile resembles an 18B model; the memory footprint to host it does not. Anyone reading "Flash" as "runs on modest hardware" has the wrong end of it — you need capacity for all 320B in memory to serve it at all.

What the common framing gets wrong

The line being quoted is that the model "outperforms GLM-5.2 across benchmarks and real-world workloads at one-tenth the price." Read the subject of that sentence. The comparison is to GLM-5.2 — Z.ai's own previous release — not to any frontier model from another lab. A tenfold price reduction against your own prior generation is a statement about your inference stack, not about your position relative to competitors. It is being repeated as though it means the model is a tenth the price of the frontier, which the card does not claim anywhere.

Who ran the evaluations, and under what conditions

The reported scores are 84.3 on Terminal-Bench 2.1, 63.4 on Deep-SWE and 55.3 on HLE with tools on the full set. The conditions are disclosed, and they are generous. Terminal-Bench 2.1 was run at temperature 1.0 with max_new_tokens of 65,536 and a six-hour timeout. Deep-SWE used a six-hour timeout and a 400K context. HLE allowed a maximum generation length of 163,840 tokens. These are self-reported figures under self-selected harnesses, which is standard practice and also the reason agentic benchmark numbers from different labs are rarely comparable to one another.

What the MIT licence buys

The licence is the part of this release that is unambiguous. MIT imposes no field-of-use restriction, no acceptable-use policy, no revenue threshold above which terms change and no obligation to publish derivatives. That is materially more permissive than the community licences attached to most open-weight releases of comparable scale. For a 320B-parameter model with a 300,000-token evaluated context, that combination — frontier-adjacent scale, fully permissive terms — is the actual news, and it is not what the price headline is drawing attention to.