GLM 5.3 is cloud-only on Ollama, and antirez is running it across two Macs
Ollama's library lists no downloadable tag for the biggest open-weights model it carries, and the people still running it at home are doing it at two bits.
The size column on Ollama's page for the largest open-weights model it carries is a dash.
Ollama finished rolling out GLM 5.3 and its Flash sibling on 29 August. Both exist in the library under exactly one tag each, and both tags end in :cloud.
There is no local build.
Not a big one, not a quantised one, not a four-bit one for people who have a great deal of unified memory. For a company whose reason to exist was ollama run on your own machine, that is a quiet and fairly large change, and it is worth being specific about why rather than treating it as a betrayal.
The arithmetic here is ours, because Ollama leaves it to you
Its own page puts the model at 753B parameters. Four bits a weight puts the file somewhere near 376 GB before a single token of cache lands anywhere. Two bits gets you to roughly 188 GB, which is still more memory than almost anybody has under a desk.
Flash is smaller, at 321B total with 18B active, which is the version a well-funded desk could actually hold. Ollama does not offer that one locally either.
So who is still running these on their own hardware?
Salvatore Sanfilippo, who wrote Redis and posts as antirez, for one. On 3 September he published an extensive benchmark of quantised local runs, a mix of cybersecurity, maths and QA tasks, twelve per model, with reruns.
The last row is the one that says where consumer inference has got to. A four-bit quantisation, run tensor-parallel across two Macs, finishing ten of twelve tasks after reruns. At two bits it managed nine.
Across everything he ran, completion rose from 43 of 72 tasks to 56 on a strict count, or 57 if you allow a DeepSeek answer he describes as "mathematically correct but incorrectly formatted."
His note on what still fails is the useful part for anyone tempted to try this. "The remaining failures are predominantly long Olympiad proofs that identify promising or correct results but never emit a final answer." Not wrong answers. No answer at all.
And one result cuts against the obvious advice. "Sampling substantially improved DeepSeek, but GLM Q2's greedy 16K run scored 10/12 versus 9/12 using normal sampling." If you are running a heavily quantised model, the sampling settings that help one can cost you on another. A nuisance, and a headline nowhere.
Ollama has become an inference reseller at the top of its catalogue
It has not said so. The pricing page is candid enough: $1.40 per million input tokens and $4.40 output for the big model, against $0.15 and $0.50 for Flash, a factor of nine either way. And the zero-data-retention, US-and-Europe hosting language is the language of a vendor, not of a runtime.
That is a perfectly reasonable business (running other people's GPUs pays better than shipping a binary). It is also a different product from the one the name sells.
Every launch command on the page carries a :cloud suffix, and the only signal that your tokens are leaving the building is that dash in the size column (plus a readme paragraph about data retention, to be fair).
The counter-argument is largely right
A model this size is not consumer hardware, no packaging decision changes that, and offering it as a cloud endpoint beside local models is more honest than pretending a laptop can hold it. Fair.
But Flash runs 18B active, and quantises to about 160 GB at four bits. Somebody could ship that, and it would fit on a machine a serious hobbyist already owns.
We would expect no local tag for the big model before the end of 2026 arrives. A downloadable four-bit build turning up in the library would prove us wrong, and we would be pleased about it.
So what would you rather have? A dash in a size column, or a 160 GB download you probably cannot run? We would take the download, and antirez has just shown what a person can do with one.
