The Vibe Father
Z.AI founder Tang Jie was asked on Weibo whether GLM still stands a chance after "epic-level" improvements from Qwen and Kimi — and he answered with two words "Epic-level plus." That's the entire announcement. No model card, no benchmark table, no date. But it's a direct reply from the person running the lab, and it landed the same day Bloomberg reported that Z.AI has finished building a 1-gigawatt data center to train its next-generation GLM models. Two words plus a gigawatt of compute is a different signal than two words alone.
The exchange was surfaced to a wider audience by @AiBattle_ on X on July 20, 2026, pulling about 14.6K views in its first hours
- Qwen and Kimi have both undergone epic-level improvements. Does GLM still stand a chance?
- Tang Jie (Founder of z.ai). Epic-level Plus.
Really excited for the next GLM model
What actually happened
The setup matters. On July 19, a Weibo user named caihzd, posting from Guangdong, left a comment under one of Tang Jie's posts "Qwen and Kimi have both undergone epic-level improvements. Does GLM still stand a chance?" Tang Jie, replying as the post's author, wrote "Epic-level plus."
That phrasing is doing real work. The commenter's framing — "epic-level improvements" — isn't an exaggeration. Qwen's 3.8 preview line has been improving at a pace we've been tracking almost daily, and Kimi K3's launch has been one of the loudest model moments of the summer (we've covered everything from the subscription sell-out to the search surges around it). Asking whether GLM can keep up is a fair question right now, not a troll.
Tang Jie's answer accepts the premise — yes, those jumps were epic-level — and claims the next GLM clears it. Founders say confident things on social media all the time. The reason this one is worth more than a shrug is what sits underneath it.
The gigawatt behind the two words
On the same day the exchange went viral, Bloomberg reported that Z.AI has completed a 1-gigawatt data center built entirely on Chinese chips, intended for training next-generation GLM models. We covered the details in our report on the gigawatt buildout — the short version is that this is one of the largest single training-infrastructure commitments any Chinese AI lab has made, and it removes the most obvious constraint on Z.AI's next run, compute.
Put the two facts side by side and "Epic-level plus" stops reading like trash talk. A founder teasing a bigger model while his company finishes a purpose-built training campus is describing a plan, not a mood. My read, the comment section bravado is real, but the data center is the actual statement. The Weibo reply just tells you the confidence level inside the building.
Worth being skeptical here on one point, a finished data center says nothing about how far along the training run itself is. Gigawatts are necessary, not sufficient. Nobody outside Z.AI knows whether the next GLM is weeks away or months away.
Why the tease rode on a sparse-attention paper
Here's the detail most recap posts skipped. Tang Jie's "Epic-level plus" reply wasn't floating in a vacuum. It was attached to his post about IndexCache — a paper from his team titled "Accelerating Sparse Attention via Cross-Layer Index Reuse," published at COLM, an international conference on large language models.
The idea, per the post, is almost embarrassingly practical. Deep Sparse Attention (DSA) makes attention faster but leaves a hidden O(L²) overhead, the indexer has to recompute which tokens to attend to at every layer. Tang Jie's team observed that adjacent layers select almost identical tokens — so most layers don't need their own index computation at all. Reuse the index across layers, and the overhead largely disappears. The kicker, according to the post, it can be implemented by changing only a few lines of code.
Now the informed speculation — and I'll label it as that. Founders don't usually post efficiency research the week before teasing a flagship model by accident. Efficiency work is what you ship when you're thinking about serving costs and long-context performance at scale. If the next GLM leans on sparse-attention efficiency in the IndexCache vein, the play would be a model that's cheaper and faster to run at long context — which happens to be exactly the profile that matters for agentic coding workloads. That's inference from the timing, not a leaked spec sheet. Treat it accordingly.
What this means for developers picking models
If you write code with AI for a living, here's why a two-word Weibo reply is worth 30 seconds of your attention. GLM has already been competitive on coding and agentic tasks. We've been watching GLM 5.2 climb alongside the Kimi K3 hype, and it earns its spot in multi-model harnesses on real repository work — not just leaderboard screenshots. A genuinely "epic-level plus" successor would land directly in the workflow of every vibe coder currently routing between Qwen, Kimi, and GLM seats.
The practical guidance doesn't change with a tease, though
- Don't pre-commit. No name, no specs, no date exist yet. Anyone selling you "GLM 5.3 early access" is inventing a product.
- Keep your harness multi-model. Qwen and Kimi are shipping real improvements today. The cost of waiting for GLM's next move is zero if your setup already routes between models.
- Watch the efficiency angle. If IndexCache-style sparse attention ships in the next GLM, long-context repo work could get meaningfully cheaper. That's the specific claim to test first.
- Benchmark it yourself when it drops. We will. As soon as the next GLM is publicly available, we'll run it through the same coding and agentic suites as everything else on The Vibe Father benchmarks — same tasks, same harness, no cherry-picked demos.
What we don't know
Let's be exact about the fact floor, because the hype cycle around Chinese labs has been running hot this month
- No model name has been confirmed. You'll see version numbers floating around social media already. None of them come from Z.AI. Tang Jie said "Epic-level plus," full stop.
- No specs. Parameter count, context window, pricing, open-weight status — all unknown.
- No date. The completed data center tells you Z.AI can train at scale, not when a release lands.
- No benchmarks. "Plus" is a marketing word until someone runs the model. Ourselves included.
What's confirmed is narrower and more interesting, the founder of Z.AI publicly committed to beating the current Qwen/Kimi generation, on the record, attached to his own name, on the same day his company's gigawatt training campus made the business press. That's a promise the whole industry can now hold him to.
What to watch, any official Z.AI channels for the actual announcement, the benchmark drops in the first 48 hours after release, and whether IndexCache-style sparse attention shows up in the technical report. When it lands, we'll have numbers on the benchmarks page — and then we find out what "plus" costs in tokens.
