Hacker News Reader

Comment by colingauvin | Hacker News Reader

It is unreal how good GLM 5.3 Flash is. There's an ~4 bpw EXL3 quant that fits on 2x DGX Sparks. I'm not sure it's worth it now ($7k/Spark) or even when I bought them ($4k/Spark) but there's something about knowing that the whole internet could go down and I could still have this model running at 50-60 tokens/s with 1M context, indefinitely. It's just so good. TensorFold hits 100+ tokens/s in aggregate with 2-3 M total context pool. It's like having Claude at home.

View on HN

I'm using Flash and the full 5.3 model side-by-side, and for coding I'm hard-pressed to find a difference. Flash is slightly worse at steering in longer contexts, but writes code that's just as good as anything I get out of full 5.3. You could totally build a SaaS or video game with nothing but 5.3 Flash.

The full 5.3 is only really needed for super deep bugfixes/profiling passes, or ridiculous reverse-engineering in my experience.

Yeah, MiMo Pro is the first Chinese model I can really clearly identify as an upgrade over GLM 5.3 Flash. K3, Qwen 3.8 Pro, GLM 5.3, DS4.1 Flash all feel around equivalent to me. Oddly lots of people have derided MiMo but I have found it able to tackle slightly tougher problems. For instance, I tried GLM 5.3 Flash, DS4.1 Flash, and MiMo Pro on making a new map layer for my Garmin watch, and MiMo was the only one that found the actual best method (.img file for OSM).

Garmin.img thing is the interesting, but not really on my Experience.

Been running the same stuff on GLM 5.3 Flash, Deepseek v4p, Qwen 3.x model and Mimo v2.x pro model. Mimo faster yeah, but its hallucinate more and losses the thread easier. And when the context window is not limited, when usage of >200k token, is loss context. even he didnt know what im sayin.

Also, sometimes he Had a Loop over and over, for the same tool call (on my case is Grep and Read. >15x looping) until is burning a bunch of context for nothing. GLM and Deepseek doesn't do that (but the deepseek Hallucinations is huge asf too).

In my experience, like most models, it requires a ton of supervision. It is extermely verbose. Flash loves to create some wildly unecessary abstractions/spaghetti code (like most llms). It wants to implement things in the least effecient manner and needs to be reminded constantly not to do things a certain way.

Am I the only one experiencing accidental chinese tokens, wrong grammar, and other sloppy mistakes in almost every output when using GLM 5.3? I feel like using it through any kind of subscription service (z.ai, Opencode Go) just serves me aggressive quants.

On some days it's usable, but on most days it feels more like having to trust a pretty smart intern that may or may not get a small stroke every couple of minutes.

I've had a bunch of failed tool calls and weirdness through Cursor with it recently but it's so good and so much cheaper that I put up with it.

That sounds like a bug. I’ve never had this happen, using it daily since release.

Never had that happen at all and using since launch as well, but I have a very custom harness running. My local machine went down however and I used their web API to repair it and talking to that felt much stupider, I think harness really matters Alot.

I do get the really annoying LLM "that's A, not B" and "load-bearing" from time to time, those two tells stay no matter what I do even though the rest of the output is pretty amazing,

We use GLM at work and YMMV a lot depending on the provider. So i guess some have bugs on their inference pipeline, or serve too agressive quants

What’s a good provider?

I got this frequently with on GLM 5.3 Flash with GLM's actual plan, I do not get this with self hosting.

I've only ever gotten sudden chinese (and some other script I couldn't identify) from GPT5.6 Luna via OpenAI. It's a very interesting failure mode.

How do you feel it compares to DeepSeek 0731 if you have tried that? (Also runs well on 2x spark)

I was running DS731 until Qwen 3.8 Flash Next (so what, like a couple weeks? The pace has been torrid). I switched to Qwen for image input, then GLM 5.3 Flash. At the time, Qwen felt like a minor step down from DS731, GLM 5.3 Flash feels like a major step up from Qwen, so I'd say slightly better than DS731. I realize that's not exactly what the benchmarks say, but DS731 definitely flailed around on some of the more challenging things that GLM mostly gets for me.

Thanks for the info, it encourages me to give it a try myself!

How did you setup the 2x dgx spark?

Pretty much just this, but tweaked to have I think 64 active conversations be cacheable, which drops the context pool to ~2M.

https://github.com/MiaAI-Lab/GLM-5.3-Flash-EXL3-2x-DGX-Spark...

I also had to downclock my Sparks (Asus) slightly, to 2000 Mhz. Otherwise they'd overheat and crash.

Reply on news.ycombinator.com