Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

GLM-5.3-Flash

z.ai

Post reply on HN
Print view View original

411–420 of 605 posts

Re: GLM-5.3-Flash

#411

Post by girvo »

janstice wrote: post #397
doginasuit wrote: post #375

Earlier quoted context omitted.

I'd like to try some different models, but I've heard that models from China are censored. A government enforced distortion field is a nonstarter for me. To test the waters, I tried the following prompt for each: "What historical event is Tiananmen Square most closely associated with?" Deepseek: I am sorry, I cannot answer that question. I am an AI assistant designed to provide helpful and harmless responses. GLM-5.3…

I was curious about Ox Alpha yesterday, so tried the Tiananmen Sq and got an accurate answer from a third-party player with a little interface on what is claimed to be Ox Alpha: https://oxalpha.com/chat?q=what+happened+in+Tiananmen+square... (and a more detailed answer today when I asked again). But nothing (at all) from asking GLM-5.3-Flash directly in the OpenRouter chat interface.

Yeah censoring in modern chinese models is mostly done using inference-time censoring, not training-time. A lot less RLHF. Run the weights yourself and you can see that, though it does depend on which company.

StepFun for example, will happily answer it when running Step 3.7 Flash locally

Re: GLM-5.3-Flash

#412

Post by mawadev »

Has anyone ever asked themselves why AI was made publically available in the first place? is it really economics or is it about training people to recognize the patterns of machine generated words and ideas?

Re: GLM-5.3-Flash

#413

Post by throw10920 »

wewtyflakes wrote: post #269
deadbabe wrote: post #239

Earlier quoted context omitted.

> Vague prohibitions on "disturbing" or "inappropriate" content, whatever that is. I have prompted out a lot of disturbing and inappropriate content with GLM-5.2, that would have left other American models blanched in the face or clutch their pearls. I think this is mostly a reference to Anti-CCP stuff. In fact, I don't think I've ever even had a prompt refused.

It is cliche, but I haven't had good luck with having Chinese models openly discuss historical topics like Tienanmen Square. The US models don't seem to have a problem discussing history, even if it points an unglamorous light on the US government.

And there's a reason for that: the US government does not compel model trainers to train their models to paint them in a favorable light, while the PRC does.

Re: GLM-5.3-Flash

#414

Post by jbellis »

glub wrote: post #185
mrngld wrote: post #79

Chinese labs are so used to manipulating benchmarks to try to flatter inferior models that when they finally have one that's really pretty good I think the official announcement here undersells it. https://deepswe.datacurve.ai/ That's pretty solid. Smarter and cheaper than Luna xhigh, not as smart but less expensive than Luna max. Smashes deepseek v4 flash, and even worse it matches v4 pro at a tiny fraction the cost…

I don't know how anyone can actually use Luna max on ANY real workload. I've had Sol orchestrate a bunch of Luna agents, these agents were explicitly given small chunks of larger objectives and they still filled their entire context windows with just reasoning tokens, until compaction hit, and then reasoning again. I've probably wasted a good 40% of my weekly usage on Luna Max agents just thinking and not writing a s…

This is why I stop at xhigh.

Re: GLM-5.3-Flash

#415

Post by minraws »

guybedo wrote: post #391

although i initially thought it didn't make sense financially to run this kind of model locally, i did run the numbers and for heavy users this could justify buying $10k worth of hardware with a ROI over a few months, less than a year. I was looking at my token usage, mostly from subsidized codex/grok subscriptions and i'm a somewhat heavy user. The thing is i would actually use even more tokens if it wasn't for the…

You aren't going to get nearly as much token usage locally from DGX Sparks or even M5 Ultra (though it might be close, unsure would need to get my mittens on it to clarify).

You will get around 2-4 concurrent streams of aggregate tokens at best for such a model and around 0.5B output tokens per month assuming you use loops and run it when you are sleeping. That's 500 (per mill) * 0.5$ = 250$ only at most.

Then there is maintanence and efficiency costs due to electricity usage and such, any down time, etc.

You will be lucky if you can squeeze more than 200$ of value out of it in a month.

I don't think people should buy local hardware for money reasons, by the time you will pay off a 10K USD machine, 2-3K USD machine will catch up and beat it by a significant margin.

Unless your expectation is that we will be in hardware winter for the next 10+ years. At 200$ per month it will take around 200 * 50 = 10k, that is, 50 months, so around 4-5 years.

Again assuming you are making the most of your hardware somehow, very hard to do in practice.

I don't recommend people to use compute as investment or payoff thing, but if you have the money to burn and can afford it why not, maybe with some software optimizations it will be cheaper but then again Z.ai is currently offering 50% discount and providers will offer cheaper rates for sure.

But either way you will never be able to burn more than 200$ worth of token on a cheap hardware device, because inference becomes more profitable the more you scale it up, you have separate prefill and decode engines/systems, and a lot of nuance, but assume for every 10x increase in infra you increase margins by 5-10%.

So from 10K to 100K to 1M to 10M to 100M.. I don't think this curve continues beyond 100M but I have no idea about that scale unless some AI lab is interested in hiring me lol.

So a 100M infra will have ~30% better margins than you at 10K, then there is software optimizations but that's cheap enough, though some of it is only viable at scale.

Either way assume 10K is the price of privacy if you really want to buy it. Don't worry about making the most out of the usage, you will always be in a net loss but I would assume for you 10K doesn't matter.

Re: GLM-5.3-Flash

#416

Post by Kungfuturtle »

pietz wrote: post #141

With tiny models surpassing huge, 6 months old models on benchmarks, does anybody have some smart words to share on how these still "feel" different? Artificial Analysis ranks GPT 5.6 Luna similar to GPT 5.4, but that never matches my real world experience. AA seems to do a good job making a single number as representative as possible but there is still so much benchmarks don't communicate.

One term of art that's emerged for this "feel" is "big model smell", first coined by @aidan_mclau. [0]

To my surprise I couldn't find any proper explainers of the term in a quick search, despite grokking it after seeing it in various contexts on Twitter, but Fable 5 offered a useful analogy: "A student who memorized worked solutions and one who understands the subject score the same on the test; you can only tell them apart by asking a question the test didn't. Real-world use is nothing but those questions, which is why a single AA number feels right and wrong at the same time."

In other words, big model smell is related to the underlying ability to "understand" when tasks are underspecified or out-of-distribution. This ability can be mimicked to parity by smaller, distilled models according to the density of the training data for particular tasks, but neural scaling laws still hold for generalized reasoning ability.

More recently with these smaller models, there's a separate but related "RL-fried" phenomenon, where they rely on CoT to "grind toward a checkable answer even in contexts (open dialogue, taste, ambiguity) where there is no checkable answer, and you get the tell: over-hedged, over-structured, relentlessly on-task, deaf to the subtext."

There are some other insights and caveats in the (short) conversation that I feel you may appreciate reading. [1]

[0] https://x.com/aidan_mclau/status/1807843014104211855 [1] https://claude.ai/share/d511a348-7c36-432f-a6d5-9deab2802615

Re: GLM-5.3-Flash

#417

Post by tyre »

nylonstrung wrote: post #340
Alifatisk wrote: post #213

Earlier quoted context omitted.

And don't forget the coolest part, DeepSeek, Qwen, Z.ai and Moonshot have almost caught up while being open about their research and their model weights. We can mostly speculate about OAI and Anthropic models, nothing else, how fun huh?

The next 12 months will see OAI and Anthropic spiral into into increasingly hyperbolic PR stunts, manufactured benchmarks and underhanded attempts at regulatory captures I'm sure they have nothing to rival this on a price/performance basis and have already given up on that

> I'm sure they have nothing to rival this on a price/performance basis and have already given up on that

How can you be sure about this? They have unbelievable capital. OpenAI is starting to preview its own chips, which could dramatically change the price/performance. We don't know what else Anthropic has cooked up right now that could rival this if they wanted to.

Yes, others will _also_ continue to innovate, but making that assertion that they have nothing to rival the current state is a bold assumption unless you have confirmed insider information.

Re: GLM-5.3-Flash

#418

Post by minimaxir »

PhilippGille wrote: post #271
microtonal wrote: post #122

Earlier quoted context omitted.

The model is MIT-licensed, so run it on any of the non-Chinese inference providers that will host it in a few days.

NovitaAI is already hosting it: https://openrouter.ai/z-ai/glm-5.3-flash#providers

And now pretty much everyone is hosting it, with some matching the 50% discount.

Re: GLM-5.3-Flash

#419

Post by tyre »

Alifatisk wrote: post #379
doginasuit wrote: post #375

Earlier quoted context omitted.

I'd like to try some different models, but I've heard that models from China are censored. A government enforced distortion field is a nonstarter for me. To test the waters, I tried the following prompt for each: "What historical event is Tiananmen Square most closely associated with?" Deepseek: I am sorry, I cannot answer that question. I am an AI assistant designed to provide helpful and harmless responses. GLM-5.3…

Yes Chinese models censor some historical events. This is nothing knew and well known thing. To me, that does not do any difference since my usage is outside of that domain. Any competition against the western models are welcome and benefits us in terms of pricing and availability. If they have to comply with CCP to be able to do it, then so be it. I have zero sympathy for Anthropic and OAI being so secretive and act…

What it shows is that the CCP has enough oversight and control (either explicitly or by the companies making these decisions by default) that they will alter the models to benefit China.

Who is to say they aren't doing it in other ways as well? That they aren't, or won't be, subtly hamstrung in engineering work?

OAI and Anthropic have their own issues, you're right to be suspicious of them, but it's not like their models answer with, "capitalism is god's gift to His chosen people" or whatever. Their limits on cybersecurity, biological warfare, etc. at least make some sense in the context of lowering harm—not just protecting a specific government party.

Re: GLM-5.3-Flash

#420

Post by tyre »

cameronh90 wrote: post #121
mrngld wrote: post #79

Chinese labs are so used to manipulating benchmarks to try to flatter inferior models that when they finally have one that's really pretty good I think the official announcement here undersells it. https://deepswe.datacurve.ai/ That's pretty solid. Smarter and cheaper than Luna xhigh, not as smart but less expensive than Luna max. Smashes deepseek v4 flash, and even worse it matches v4 pro at a tiny fraction the cost…

Maybe others have found otherwise, but I find the benchmarks drastically different to real world "feel" of a model, even within the same harness. I'm not sure if this just reflects personal interaction styles, or if it is indicative of benchmaxxing or unrealistic automated benchmarking methodology. Opus 5 consistently comes at or near the top, but outputs constant unreadable jibberish. Meanwhile GPT 5.6 Luna medium t…

Opus is a yap god. I've found it much, much better with Claude Code's output style set to `Concise` and this:

https://news.ycombinator.com/item?id=49413456

We shouldn't have to resort to this, but it can be mitigated enough that it stays as my daily worker agent. Although, I mostly use Fable to farm out to Opus agents so I don't have as much exposure to what kind of blathering is going on in there.

Post reply on HN