Looks alright for a "first" but there's no reason for anyone to really use until they open source it.

it's an agent product that you interact with over email. it has skills and stuff. it has multiplayer. so it's something like openclaw, but different.

you can counter the context rot and requirement drift that is experienced here by many users by using a recursive, self-documenting workflow: https://github.com/doubleuuser/rlm-workflow

you can use lurkkit.com to build your own chronological youtube feed with only your subscriptions

> Traditional Chinese relies on context: “Rain heavy, not go”, “雨大，不去了”. > Modern Chinese demands explicit logic: “Because the rain is heavy, therefore I will not go.””因为雨下得很大，所以我决定不去了。”

I would say "下雨了，我不去“ or something like that. The second example is perhaps what a language learner would say in order to "speak correctly", but nobody actually speaks or writes like that.

Totally. I also feel such a disconnect with HSK material, no one speaks like that or even uses that vocabulary. But I guess thats the case with almost every language/language course.

What's gone unnoticed with the Gemma 4 release is that it crowned Qwen as the small model SOTA. So for the first time a Chinese lab holds the frontier in a model category. It is a minor DeepSeek model, because western labs have to catch up with Alibaba now.

on my 16 GB GPU Gemma 4 is better and faster than Qwen 3.5, both at 4-bit

so it's not so clear cut

depends on usage, Gemma 4 is better on visuals/html/css and language understanding (Which probably plays a role in prompting). But it's worse at code in general compared to Qwen 3.5 27B.

Which in the series specifically?

It's unnoticed because it didn't. In Google's own benchmarks they are on par, and I've seen 3rd party benchmarks where Qwen beats G4 with high margin

most codebases dont have traces to train on. if you use rlm-workflow you will build up rich traceability in the form of requirements, plans, implementation artifacts, along with worktree diffs. with these, you can then use self-distillation on models or use autoagent to improve your harness. https://github.com/doubleuuser/rlm-workflow

For anyone that believes Chinese labs will stop open sourcing their models, let me tell you why that won't happen.

First, try signing up for Z.ai's coding plan. I know how to but I bet you won't be able to.

The absolute disaster that is Z.ai's internet presence shows that these small labs have no ability to market themselves and drive direct sales.

For marketing, they lack capabilities, and releasing open models is the only way for them to remain in the conversation.

For sales, they rely on distribution via OpenRouter, OpenCode etc. Interest with their users is driven by open model performance.

Open sourcing for Chinese labs is not some large national scheme. It is their only way to commercialization.

Well, can’t they just direct their model to do some marketing for them?

Only partially tongue in cheek - if it’s not good at marketing itself, that seems like a red flag for capabilities?

> First, try signing up for Z.ai's coding plan. I know how to but I bet you won't be able to.

What's the issue with signup up for Z.ais coding plan?

you can't find it

1. Go to https://z.ai

2. Top right corner, "API"

3. https://z.ai/subscribe

China can't get good chips. But I don't understand why they can't license their closed source models to US inference providers so we can get more than 80% reliability on their models on OpenRouter.

I think they already are licensing their biggest models to third party inference providers.

Agree, it’s surprising that these companies don’t thoroughly test their own workflows to ensure a smooth and seamless user experience.

The biggest story here is that this is Google handing Qwen the SOTA crown for small and medium models.

For the first time ever, a Chinese lab is at the frontier. Google and Nvidia are significantly behind, not just on benchmarks but real-world performance like tool calling accuracy.