In this video, I look at GLM 5.3 flash from ZAI. We look at what it can do, how it was made and how it compares to its bigger sibling GLM 5.3.
📖 Blog:
📖 Blog:
🤗 HF:
Twitter:
🕵️ Interested in building LLM Agents? Fill out the form below
Building LLM Agents Form:
👨💻Github:
⏱️Time Stamps:
00:00 Intro
00:56 GLM 5.3 Blog
01:35 GLM 5.2 Blog
01:43 GLM 5.3 Benchmarks
01:51 GLM 5.3-Flash Blog
02:04 GLM 5.3 vs GLM 5.3-Flash Comparison
03:49 GLM 5.3-Flash Price
05:17 Architecture
05:53 Benchmarks
07:44 Artificial Analysis
09:27 Demo
source
Have you noticed the awful speed of this 5.3 Flash which is supposed to be flashy? It is twice worse than even the full version of GLM 5.3 and 7x worse than Google Gemini Flash 3.7. it is worst of the worst off all AI models in the market. What's the point to use it besides it's working with images now.
This is why Jensen Huang was complaining about that China will build their own chips to run these models anyway,regardless of the ban to China. I would love to see these to go main-stream with a mess production,so local AI users shall benefit.Maybe in 2 years of time running these models locally with an "affordable" price,will not be a dream anymore. Nice video Sam.Thanks a lot.
My dude go to speech therapy please
"If somebody does fix the long thinking, [qwen 3.8 27b] will be a killer local model for a long time to come"
Yea, for about a month until v4.0 is released. 😉
This model is actually insane, and it has replaced DeepSeek fully for me.
Unfortunately, from the provider, GLM-5.3 flash is painfully slow. Regular GLM-5.3 is 3x faster and uses less tokens. Unless I am using the vision capabilities, definitely sticking with the original.
I tried using it as Ox Alpha, and it made a huge mess. GLM 5.3 for planning/review + DeepSeek V4 Flash (low) for build has been a solid combo, but Ox Alpha consistently failed at all three: it failed to properly gather necessary context for creating plans or would disregard important details when making plans; it failed to complete the full scope of tasks, constantly leaving things undone or only half-done; it failed to catch mistakes during review. Idk what everyone else is seeing in this model. DeepSeek V4 Flash (low) has been far superior in my experience working in a real software project. But then again, I am developing libraries that require precision, so perhaps it is a perfect fit for people who just want to slop it up.
This is not about Google secretly breaking Bitcoin today. The real story is that the estimated cost of a future quantum attack may be falling faster than expected.
Google's researchers found ways to reduce the resources a future fault-tolerant quantum computer could need to attack 256-bit elliptic-curve cryptography.
But instead of publishing the complete optimized circuit, they used a zero-knowledge proof so others could verify the central result without receiving every detail needed to reproduce the attack.
That's a pretty unusual trade-off:
Scientific transparency cybersecurity
Normally, researchers publish everything so others can reproduce the work. But with cryptography, publishing the complete method could potentially give attackers a roadmap before vulnerable systems have migrated to post-quantum encryption.
The criticism is still fair:
If we can't inspect the complete circuit, how much are we independently verifying—and how much are we trusting the researchers?
The important point is that current quantum computers cannot perform this attack.
But the direction matters:
better quantum algorithms → fewer resources required → lower future attack cost → greater pressure to migrate to post-quantum cryptography.
So the real race isn't:
“Can quantum computers break encryption today?”
It's:
Can the world replace vulnerable cryptography before quantum hardware and algorithms become powerful enough to attack it?
That's why this research matters. The threat is still future-facing—but the window for preparing for it is getting smalle
Awesome work Sam! Can't wait to see if this is usable on a 128GB unified ram system… 😅
I’ve been using it on Cloudflare and open router and the way it does the thinking it’s actually quite good if you want to watch it because it’s it’s readable and you can see it’s working on and a summary at the end.
What I love about opencode is that I can have sol use GLM-5.3 flash sub-agent
I swapped in my internal workloads and the savings are enormous indeed specially with the half price discount
But the flash version is a bit more verbose
9:20 … A issue with multiple benchmarks is that they absolute refuse to test Chinese models at any thinking effort lower then Max. GLM 5.3 Flash at High gives you almost the same intelligence but at ~2/5 output tokens. Something you normally only see on Frontier level models, where the gap between Max and High, is often not so big because of the models massive parameter training.
So which thinking level is recommended to be used in general?
You know a model is good when they have to tease it as "Ox-alpha" and not just the model itself
9:10 That's GLM 5.3 full..
It is best intelligence per dollar I feel
I really like this new style of video that is not focused solely on benchmarks or feeling (still relevant as a first shot), continue like this 👏🏻
❤❤❤
“OKIIIIIIIIIII”
Thanks for the overview, Sam. It is getting to the point where I don’t consider a model “released” until I’ve gotten the Witteveen Overview 😁. I must say, though, I am really getting benchmark fatigue. Every time you turn around, there is a new benchmark, everyone chooses who to compare with, that proves their point or shows off their model, and almost every graph has a different x and y axis. 😮
I am curious, Sam, what you did with your 100 million tokens on Qwen 3.8 27B ?
you made so much sponsored content that I can't even appreciate your non sponsored content anymore
Screw benchmarks, 5.3 flash is better than Sol on my tests, and definitely better than the other open weights.
everyone compares token price, but ignores retry rates and failed tool calls; a cheaper model gets expensive fast when agents need three attempts per task
GLM-5.3 isn’t usable on their subscription model for me because I get stopped out about halfway through their 5-hr token limit repeatedly.