GLM-5.3: Frontier coding with emergent cyber capabilities

z.ai403 pointsby pella166 comments

Join the discussion

Write your take first — we'll ask for email only when you're ready to publish.

Hacker News
Their coding plan switched to credits, didn’t it? What are the rate limits like, compared to Anthropic or Kimi K3?
I remember trying their Coding Plan out before the change and the 5 hour limits felt too restrictive then even for light/medium work, especially cause of the whole peak and off-peak thing: https://blog.kronis.dev/blog/z-ai-s-glm-5-2-is-a-great-model...
Nowadays, I’d probably go with their Max plan if the rate limits are okay? Anyone using them now?
Oh also unrelated but ZCode was surprisingly good, which is surprising for a tool that came out of nowhere - even some of the critiques in my blog post have been patched out. Sadly they don’t support using Claude Code as an agent so can’t use it like Paseo or Kepler or Agent Orchestrator.
by KronisLV
I feel like quota on their subs is extremely generous. I pay 3-4 times less for larger quota than gpt-5.6-sol.
by scotty79
I found with GLM I was better off using plans from either Neuralwatt or Ollama.
But Neuralwatt significantly raised their rates since then.
by cmrdporcupine
Update: tested it out myself on their Max plan, on some parallel agentic sessions.
Currently 20% of my 5 hour limit and 4% of my weekly limit.
```
  Total: 58.46M
  GLM-5.3 Cached: 56.91M
  GLM-5.3 Uncached: 1.23M
  GLM-5.3 Output: 315.18K
  Cache hit rate: 97.9%
```
Extrapolating from that (inaccurate for now but oh well):
```
            Full 5-hour  Full weekly
  Total     292.3M       1.461B
  Cached    284.6M       1.423B
  Uncached  6.15M        30.75M
  Output    1.576M       7.88M
```
All of the work was off-peak I think, using OpenCode not ZCode in these examples.
Their own estimates are quite different, probably due to their conservative caching estimates vs what I normally get on longer form work: https://docs.z.ai/devpack/overview#estimated-token-allowance
by KronisLV
I used the legacy plan with only the 5 hour limits on lite. It was super awesome but sadly not comparable to the new plan.
Currently I am on the new max plan with the 5hr limits and weekly limits, I can't speak for the credit plan. Using it exclusively in zcode because of the usage multiplier + the harness is genuinely good.
Easily do a billion tokens per week on my limits and generally have no problems with limits, however if I use it during peak hours I will hit the 5hr limit super fast even in zcode.
Zcode gives much more usage: Normal hours 1x -> 0.67x usage multiplier Peak hours 3x -> 2x usage multiplier
Peak lines up with the afternoon for me and I prefer coding morning/night so its not really a problem + I have the codex $20 plan and opencode go so I can always use other subs during peak hours.
If the peak hours are your main work hours (its a 4 hour window) then the value is going to be MUCH lower, especially in CC or other harnesses (1/3rd the usage limits is harsh).
It does also change a bit depending on demand, so I recommend getting the max plan because you get priority access if you really like the model, it would be a 10/10 recommendation for me if it had vision but rn its mainly useful for backend or throw away internal tools where IDK about the UI as much.
by ryan-a
Wdym "sadly they don’t support using Claude Code"? For the longest time that's all Zai supported - Claude code. I'd run it via
```
  export ZAI_ANTHROPIC_BASE_URL="https://api.z.ai/api/anthropic"
  export ZAI_ANTHROPIC_AUTH_TOKEN="$ZAI_API_KEY"
  claude-zai() {
      { local -; set -x; } 2>/dev/null
      ANTHROPIC_BASE_URL="$ZAI_ANTHROPIC_BASE_URL" ANTHROPIC_AUTH_TOKEN="$ZAI_ANTHROPIC_AUTH_TOKEN" claude "$@"
  }
  $ claude-zai
```
I liked Claude Code to start with. But over time between 'CC cache thrashing undo' seetings (I see now accumulated in ~/.claude/settings.json) and Anthropic-anything becoming a liability - have not used it in while. ZCode is ok and use it to take advantage of the discount tokens on offer from time to time. But really glad to see that in omp (oh-my-pi) Zai is a 1st class provider, can be selected on it's own no configs shananigans needed. And fits in the overall picture. E.g. can select GLM-5.2 (now 5.3) assign role [plan] or glm-5-turbo [advisor].
Got reminded now of glm-5v-turbo - that 'v' was for vision - will try assign it role [vision] now in omp. See what happens. :-) Often times it's handy when describing gui problems if the harness/model 'can see'.
by ljosifov
> Anyone using them now?
You're gonna have a had time getting straight answer to that out of the internet. There are now 4 different flavours of the Max plan floating around (Legacy V1, Legacy V2, New plans, and the current credit ones). And on top of that they have peak times. So ~8 scenarios, 24 in total across all feedback for their coding plans.
So when someone tells you they're having a good time on a GLM coding plan it's damn near unusable as a datapoint unless both parties are very clear about what precisely is being discussed
[It's been good for me though...V1 Max off peak...which is basically the best of the 24]
by Havoc
Missing multimodal again?
It is so valuable in practise to be able to have the models see screenshots - I guess if they aren't in the benchmarks then nobody will focus on it. But it completely nixes these for some of my main use cases.
by zmmmmm
I would assume that GLM 6 will be multimodal, but 5.x will be text-only.
by arcanemachiner
I can’t come up with a use case where I couldn’t extract the image details using another, multimodal model and pass it into the GLM’s context with as many details as I need.
by pllbnk
Probably not what you're after, but I've considered having a separate small mm-model act as a seeing-eye dog for the bigger more capable one.
by xscott
Apple will release M7 MacBook Pros / Mac Minis next year, and they will be able to run free LLMs locally at native speed. All software developer notebooks will be replaced to run local models, saving a lot by cancelling Claude Code subscriptions. Developers win. Apple stocks will be rocketing. Everything else will go down. You're welcome.
by vmware508
Every single MacBook built in the past half-decade already has an LLM built into the latest version of their OS.
But there's a significant difference in hardware required between running a 3B parameter model and a 700B-1T+ parameter model.
by kube-system
Sure buddy, all you'll end up with is a $10k machine that run gimped models at like 30tok/s for about 5m before the fan kicks in and it starts to sound like a turboprop, while offering maybe 30% of the context size of hosted models.
by toasty228
> run free LLMs locally at native speed
This reads like a hallucination. What does native speed even mean?
by Flavius
I don't know why you'd want to burden your laptop with a large model. But I can totally see a new "developer workstation" product that's just a semi-large box that's optimized for running frontier open weights models for one to few users.
by scotty79
They will cost an insane amount as well. Maybe less than subscriptions or tokens. But running massive models on laptops with batteries and poor cooling doesn’t make much sense.
by Gecko4072
Local vs remote compute is a constant thread in tech history - mainframes and desktops then local and cloud compute (think Google Photos bs Apple photos - one indexes on device the other indexes in cloud). Now we have the next chapter local vs cloud LLM models.
There will always be a market for frontier labs in the cloud based models - these models will always be able to be bigger, and that will likely translate to doing things local models can’t.
Logically also we’ll likely get to a point where RAM drops in price as production ramps up, and local LLM is both capable and cost effective. This feels like it is coming for Siri / Gemini / Alexa personal assistant type use cases.
So I think the local LLM will become a thing in laptops and phones in a year or two, offering PA type use cases. Professional LLM services will likely remain at the frontier (and in the cloud) for the foreseeable.
by gehsty
That’s not how that works. The hosted models don’t stay still in size and capability while Apple advances. Both will advance their frontier and there will still be a gap and developers will still prefer the stronger option.
by Havoc
You'd need the 256 gb memory model which will be expensive because apple has trouble getting capacity (got turned down by cxmt). And even then you can only run a 2 bit quant which is noticeably worse than 8 bit
by schleck8
OpenAI and Anthropic need to just go ahead and give people access to the cyber models.
Otherwise we have a world of attackers using open and closed source models against a much smaller group of maintainers that are likely heavily dependent on Anthropic and OpenAI and for whom it may not be a simple matter to just get approval to start using the open model flavor of the month.
by virgildotcodes
Can't the maintainers use the same models as the attackers?
The maintainers don't need approval to use GLM.
by surgical_fire
They won't. People will notice that the models are overhyped once they can test them.
by 35129ab
At least OpenAI seems to want to do that, but the US is now forcing them to go through approvals. Anthropic seems much more hesitant.
by mindwok
Not only attackers. I have to switch to Kimi or GLM even in cases of basic issue triage on my own projects! Current guardrails are ridiculous.
by LeonidBugaev
This will be roughly on pair with Kimi K3, but using a third of its parameters.
Just 4 weeks ago the "Kimi K3 moment" was seen as a threat to Closed AI and in less than a month Z.ai have cut the parameter/RAM barrier to a third.
Congratulation to Z.ai and all the hard working Chinese researchers who are quitely boiling the frog.
by bertili
Kimi is a great model but it was clear from the start they achieved they brute forced that performance through scaling. The frontier models K3 compares to are rumoured to be smaller also. GLM on the other hand is way ahead in perf/parm but severly compute bound. Now once GLM can scale up or Kimi optimizes the training more, that gonna be fun times.
by rammler
Congrats def in order but as usual the proof will be in the pudding of actually running the thing.
GLM 5.2 has token efficiency problems. It's not a stupid model, but it takes a lot of "thinking" to produce not-stupid results. ("But wait...").
Which makes its pricing deceptive.
I tried to get by through the month of June on just GLM 5.2 and it was ... fine-ish for about two weeks. But the provider situation wasn't ideal.
by cmrdporcupine
> Scaling post-training is all we did for GLM-5.3.
Love this opening line. And wow, great results.
> As agent capability improves, much of the difficulty in scaling post-training moves from the model to the environment.
by wxw
does this suggest 5.3 is the same # of parameters as 5.2?
by tjwebbnorfolk