GLM-5.2: The Most Powerful Open Model yet and the Brutal Reality of Running It

walrus01

Before people go and drop a gargantuan sum of money on a server capable of running it entirely in GPU, there's still a fair amount of used x86-64 servers capable of running it in CPU and RAM (using llama-server) for probably under $6000. For example a Dell R640 with two older Xeon 18-core CPUs and 1TB of RAM. Test it out at a slow token/sec rate and see if it fits your needs. Same idea for Kimi.

tfirst

If model performance continues to scale with model size, I have a hard time seeing how local models will have any chance of competing with models hosted on datacenter hardware. 1. There are strong economies of scale in hosting inference (batched prompts, high uptime, shared infrastructure). 2. There are physical limits on how much memory we will be able to produce over the next few years. Demand will probably scale at least as fast as production does, so we won't be saved by falling prices.

kristianp

Irritating LLMisms: - "real architecture trick" - "the honest hardware reality of running it at home." - "What it is — and what Z.ai claims" - "The one genuinely new idea" And many more.

KaoruAoiShiho

Terrible zero value article, I am extremely surprised it is upvoted. That being said Artificial Analysis just came out with a brand new benchmark where it scored between opus 4.8 and gpt-5.5 and well behind fable-5 so it's definitely frontier-ish https://x.com/ArtificialAnlys/status/2067744637155226101

CorpOverreach

I do think it's going to get harder and harder to run bleeding-edge models; this is just the start of it. It being hard for the average joe to run these at its fullest potential is unfortunate, but the important part is that _you can_ assuming you can acquire the resources. I think that's going to be important for the sake of preserving privacy and freedom of information in the long run. We're seeing this play out right now with Anthropic originally playing the "safety" card for why they can't let everyone at Mythos and subsequently got on the US Gov't radar with access to Fable being pulled. The next biggest milestone will be an open-weights challenger to Mythos. There'll be consequences to that, but I feel those are less worse than someone else deciding what you can and can't use a model for.

lamida

Pretty sure the article is fully written by LLM without editing at all. See all the — emdash sprinkled all over.

blackoil

I think people overrate 'local' part of open Models vs private. With OpenAI my choice is 1. I have to use them, even if they decide to double the cost or work with govt to blow my country. My $5 server can't run GLM but I have choice from many providers based on my requirements of cost, data residency, political alignment.

easygenes

Article reads as though written by someone who doesn't have much experience with deployments like this. Underestimates the memory needed to run with a reasonable amount of context. Misses two other obvious targets: 1) 4x DGX Spark (or equivalent other GB10 boxes) with a switch (MikroTik CRS504 or CRS804) and TP=4. 2) 4x RTX PRO 6000 box. Probably the most practical for cost/perf if you want on-prem as an individual. Both would be best to run a 2-bit quant so everything can stay resident (article claims you could run a 4-bit quant with 4x RTX 6000 Ada, and while technically true it would mean a lot of the weights are streaming from DRAM, so it would be slow and impractical. You would need 8x RTX PRO 6000 to run 4 bit at a good speed). This model quantizes unusually well: https://unsloth.ai/docs/models/glm-5.2#quantization-analysis

ma2kx

Thats just stupid. - Why should I run it on local hardware when there are already about a dozen US provider available? - To compare the token usage per task with GLM 5.1 is worthless when GLM 5.1 is unable to do the task. - Not even z.ai itself runs the model with BF16 weights. - I couldn't care less how good the model is at drawing a pelican on a bicycle.