Vector Lab
VECTOR LAB

EST. 2025

WEEKLY UPDATE2026
BY ANDREW MEAD

GLM 5.3 Flash

OpenAI Divorces Cursor, Z.ai releases a strong small model, and realtime high quality video generation is here

Read in

News

OpenAI Divorces Cursor

OpenAI announced this week that they would be ending their partnership with Cursor, which will cause Cursor users to lose access to OpenAI models. The partnership is expected to end on November 12 of this year.

Important background for those that don’t know: Cursor was bought by SpaceX, with the deal being finalized 2 weeks ago.

According to OpenAI, this is because of a lack of trust from Elon Musk’s companies:

We are making this choice because we cannot be confident that SpaceX will use our technology within our terms of service, based on our experience with Elon Musk’s companies violating contracts.

I believe the underlying reasons here are two-fold; the first is Sam Altman and Elon Musk’s hatred of each other. They recently had a very public court battle (nothing really came of it), stemming from OpenAI’s restructuring from a non-profit to a for-profit company, which Elon (who is one of the OpenAI founders, but left in 2018) took exception to.

The second are more important reason is that OpenAI seems to be implying that SpaceXAI is distilling their models through Cursor and other harnesses to train Grok. This kind of distillation is useful, but not sufficient, to train a frontier model. Historically, it has been Anthropic that has cared about this (mostly in regards to Chinese AI Labs), but now OpenAI seems to care as well, but for domestic competition instead.

Interestingly Anthropic doubled down on their commitment to Cursor, saying that they will continue to increase compute to support Claude in Cursor. This friendliness is most likely not due to Anthopic’s love of Cursor, but rather due to the fact that they have a $45 billion dollar compute contract with SpaceX.

Cursor has said that they are speaking with the OpenAI team right now to try and resolve this, highlighting that OpenAI only serves 5% of Cursor traffic.

If you are a Cursor user and use OpenAI models, you do not need to jump ship yet, but I would be mentally prepared to do so in the coming months (you get much better value from an OpenAI subscription plan than a Cursor plan anyway, so I would recommend switch just for that reason).

Releases

GLM 5.3 Flash

There has been quite a bit of hype around an anonymous model called Ox Alpha the last few weeks, and it has been revealed that it was GLM 5.3 Flash that was being tested.

GLM 5.3 flash benchmarks

**

Despite being a “smaller” flash model that is 3x smaller (320B-A18B) than its big brother it is a very capable model. It passes the real world vibe check, and seems to be a step above DeepSeek V4 Flash and GPT 5.6 Luna, which are the top “small” models that we have right now.

Usually with these smaller models, they use a large number of tokens, partially offsetting their lower pricing, but Z.ai has done a good job restraining the reasoning length for GLM 5.3 Flash. It is still not at the same level of efficient token use as Luna, but makes up for it with increased intelligence.

Model$ per million (input)$ per million (output)Tokens per second
GLM 5.3 Flash$0.15$0.5028
DeepSeek V4 Flash$0.44$1.3282
GPT 5.6 Luna$1$684
Grok 4.6$2$657
GLM 5.2$1.40$4.4047

The model also has vision capabilities, but they are not in any way impressive, and are not at the same level as its text capabilities.

An interesting thing to note is that the model is being served by Z.ai using only Chinese GPUs, which is the first time that a top tier Chinese lab has (publicly) done so. It seems that Nvidia is starting to have some major competition from China, which should hopefully be good for GPU prices in the future.

The model is open sourced (along with the big GLM 5.3 model as well), but is a bit too large to run on non-datacenter hardware.

This is yet another very strong, smart, cost efficient model that I would recommend trying out. The best ways to get access to it are OpenCode Go or the Z.ai coding plan.

Quick Hits

Qwen 3.8 Flash

Qwen 3.8 Flash is a peek into the future of what the Qwen 4 model architecture will look like.

Qwen 3.8 flash benchmarks

They are using an Engram style architecture. This means that instead of having a big transformer model that must store information and facts alongside learning proper reasoning, there are two separate modules. The first is the usual transformer we know and love, and then a second engram “memory” module that acts as a fact lookup table (kind of like a database for the LLM’s brain).

The engram module requires very little compute, and just needs memory to load itself, so the model can effectively be split into two: the transformer module gets loaded on to a GPU, and the engram module will reside in system (CPU) memory. This will allow you to run larger models locally, as you won’t need GPUs with lots of memory.

This release is meant primarily to give the community time to implement and optimize the model in different LLM inference engines, but it is also probably the best model you can run right now on a DGX Spark, although you should be ready for some bugs if you do so.

H3 Max

MiniMax H3 has risen to the top one of the best video generation models, and it is also open source.

One of the biggest image and video inference providers Fal.ai took the model and finetuned it for even more quality and also speed, producing H3 Max.

H3 Max benchmark table

This results in the best video generation model that we have seen, and it is capable of generating video faster than realtime: it takes 3 seconds to generate a 5 second 768p video, meaning that you can have a continuous video stream from it, for a very low cost ($0.20 per 5 second video).

You can try it for free right now on Fal’s website.

On Device AI Benchmark

I have been asked many times before what the best model to run on a phone is, and now we have a leaderboard made to answer this specific question from Artificial Analysis.

Mobile device leaderboard

The benchmark aggregates scores from basic tool calling, hallucination, and reasoning benchmarks to get a final score, and also measures token usage and end to end speed on modern phones.

To answer the question of what is the best phone model right now, LFM2.5 2.6B seems to be the answer. It is near the top on all of the benchmarks, and strikes the right balance of speed, low hallucination rate, and intelligence.

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.

Stay Updated

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

← BACK TO NEWS