Vector Lab
VECTOR LAB

EST. 2025

WEEKLY UPDATE2026
BY ANDREW MEAD

GLM 5.3

GLM 5.3 becomes a top tier model and DeepSeek can see now

Read in

Releases

GLM 5.3

The Z.ai team has released an update to their increasingly popular GLM 5.2 model.

Good benchmarks

GLM 5.3 builds on its predecessor. GLM 5.2 was a model that was good for small to medium sized tasks, and fit best as a sub agent that your smarter agent could go and use.

GLM 5.3 no longer feels like a model that is a subagent, but is able to hold its own now independent of other models (or humans) baby sitting it.

This is corroborated by benchmarks like Artificial Analysis, which place it right next to Kimi K3.

I would place this model at the bottom of the top tier of models, alongside GPT 5.6 Sol, Fable, and Kimi K3.

What is notable compared to these competitors is its price. Unlike Kimi which saw its prices dramatically increase when K3 was released, Z.ai has kept their pricing the same between versions. The model does chew through tokens far more than its Chinese counterpart, but its still comes out being 25% cheaper than Kimi K3 and more ethan 4x cheaper than Fable, but is unable to pass GPT 5.6 in the real world cost category.

Model$ per million (input)$ per million (output)Tokens per second
DeepSeek V4 Flash$0.44$1.3282
Claude Fable 5$10$5040
GPT 5.6 Sol$5$3045
Kimi K3$3$1527
GLM 5.3$1.40$4.4050

The model has not been open sourced, which is a departure from previous Z.ai releases. This may be due to safety concerns (they initially released their model via their coding plan for a week, before rolling out to their API this week, to allow them to see how people use it for cybersecurity).

If you are in the enterprise space and cannot use models directly from China, this will mean you most likely will not be able to use this model.

If you don’t have any qualms with Chinese LLM providers, then I would definitely check this model out.

Quick Hits

DeepSeek V4 Flash Vision Exp

DeepSeek V4 Flash is one of the best cost effective LLMs right now, but it has been lacking multimodal capabilities.

The DeepSeek team knew this was a shortfall, and have released a Preview version of the Flash model with vision.

Benchmarks for V4 Flash Vision

The model seems to be a drop in replacement for the previous flash model.

Its also interesting to note that DeepSeek uses the Experimental and Preview naming to mean “unfinished”, (the original V4 preview models were nowhere near their current counterparts) so expect an updated version in the future with much improved capabilities.

DJ Sol

Someone had GPT 5.6 Sol (DJ Sol) put together a list of “the best” songs. Give it a listen on Spotify to see if the model has good taste.

DJ Sol playlist preview

The playlist is over 2,000 songs long, and includes 63 BrockHampton songs. I will let you decide if this is a good thing or not.

Finish

I hope you enjoyed the news this week. If you want to get the news every week, be sure to join our mailing list below.

BingDing Wood Klin Factory, Jingdezhen 1 BingDing Wood Klin Factory, Jingdezhen 2 BingDing Wood Klin Factory, Jingdezhen 3 BingDing Wood Klin Factory, Jingdezhen 4

BingDing Wood Klin Factory, Jingdezhen from Average Vermeer Enjoyer on Twitter

Stay Updated

Subscribe to get the latest AI news in your inbox every week!

← BACK TO NEWS