Zhipu is done letting July belong to its rivals. Founder Tang Jie has publicly confirmed the next flagship — GLM-5.5 — promising an upgrade he rates "epic, plus" in response to a community jab that Qwen and Kimi had both leveled up while GLM sat quiet. Reports put the release window in August 2026 and the expected scale above a trillion parameters. Behind it sits the week's real news: a ~1-gigawatt data center built entirely on Chinese chips, completed and partially operating — and a stock that crashed roughly 70% from its highs, then snapped back hard the day the data center was revealed. This page sorts what is confirmed from what is expected, as of July 23, and updates as the real numbers land.
Confirmed, reported, unknown
Confirmed:
- GLM-5.5 exists and is next. Tang Jie confirmed the model publicly, with the "epic-plus" framing — his answer to "Qwen and Kimi both made epic jumps; is GLM still in the game?" The model had circulated in rumors as "GLM-5.3" before the naming settled.
- The exchange with Musk happened. After Elon Musk suggested Chinese models still need time — allowing GLM might reach top-tier benchmark levels "by year-end" — Tang Jie's public reply was, in effect: *it won't take that long.*
- The gigawatt of domestic silicon is real. Zhipu has completed a data center drawing on the order of 1GW — power for roughly 750,000 homes — built entirely on Chinese-made AI chips, with multiple 10,000-chip clusters and partial operations already running. Bloomberg and others covered it on July 20. The facility exists to feed the GLM family; export controls, ironically, are what made an all-domestic build the rational choice.
- The stock round-trip is real too. Zhipu's Hong Kong-listed shares fell roughly 70% from their GLM-5.2-era highs, including a brutal two-day drop of over 40% — then rebounded around 27% on July 21, the session after the data center reveal. Community consensus blamed the crash more on unwound speculation and lock-up expiries than on Kimi K3 — and the rebound suggests the market agrees the compute story changes the equation.
Reported (not yet Zhipu's official spec):
- An August 2026 release window.
- Parameters above 1 trillion, with analyst speculation reaching the 1.6T class — a big step from GLM-5.2's 744B (about 40B active). Until a model card exists, every parameter figure is provisional.
Unknown:
- Benchmarks, pricing, context window — nothing yet.
- Open weights. GLM-5.2 shipped MIT-licensed, and openness has been Zhipu's signature. But no commitment has been made for 5.5, and July's pattern — K3's weights promised "by July 27", Qwen3.8's "soon" — has made launch-day openness a promise rather than a default. Whether Zhipu breaks that pattern is worth watching precisely because MIT weights are its brand.
The question that actually matters: does GLM stay cheap?
Here is the tension nobody at Zhipu has to answer until launch day. GLM-5.2's entire strategic identity is value: $1.40/$4.40 per million tokens for coding within 2.5 points of OpenAI's flagship — the best capability-per-dollar in the lineup. But July taught a clear lesson about what happens when Chinese labs ship trillion-class flagships: Kimi K3 launched at $3/$15 — the priciest Chinese-lab model ever — and independent testing then measured its real agentic cost at $10.57 per task, more than Claude Opus 4.8. Scale costs money, and "the end of super-cheap Chinese AI" is already a headline narrative.
So GLM-5.5 lands at a fork: follow Moonshot upmarket into premium flagship pricing, or use that gigawatt of self-owned, domestically-sourced compute to do what nobody else structurally can — serve a trillion-parameter model at value prices. Owning the power plant and paying no Nvidia margin is exactly the cost structure that would make the second path possible. Which fork Zhipu picks will say more about the next six months of the Chinese model market than any benchmark.
Where this fits in July's arms race
GLM-5.5 makes it four for four: every major Chinese lab has now shipped or teased a flagship inside five weeks — Kimi K3 (July 16, 2.8T), Qwen3.8 (July 19, 2.4T claimed), DeepSeek V4's stable release slipping but imminent, and now GLM-5.5 for August. The competitive clock is compressing: GLM-5.2 was the open-weight king for barely a month before K3 took the capability crown. If the August window holds, GLM-5.5 arrives just as K3's weights (and possibly V4's stable build) hit the market — the densest release quarter Chinese AI has ever had.
What to run today
GLM-5.5 is a tease; GLM-5.2 is the shipping product — MIT-licensed, 62.1 on SWE-bench Pro, $1.40/$4.40, and still the value default this guide has recommended since June (full case in the deep dive). It runs today on Turiloop behind one OpenAI-compatible key, next to DeepSeek V4, Kimi and Qwen — international card, no Chinese phone number. When GLM-5.5's API opens, it joins the lineup and the pricing calculator the same day; switching will be a model-id change.
FAQ
What is GLM-5.5? Zhipu's confirmed next flagship model, publicly teased by founder Tang Jie as an "epic-plus" upgrade over GLM-5.2. Reports point to an August 2026 release and a parameter count above 1 trillion. Official specs, benchmarks and pricing have not been published.
When is GLM-5.5 coming out? August 2026 per current reporting; Zhipu has not committed to a public date. This page updates when it does.
Will GLM-5.5 be open weights? Unknown. GLM-5.2 shipped under MIT and openness is Zhipu's calling card, but no commitment exists for 5.5 yet — and recent Chinese flagship launches have made "weights later" the pattern rather than the exception.
Why did Zhipu's stock crash? Shares fell roughly 70% from their GLM-5.2-era peak, with community analysis pointing at unwound speculation and lock-up expiries more than competition from Kimi K3. They rebounded around 27% on July 21 after Zhipu revealed a completed 1GW data center running entirely on Chinese chips.
Should I wait for GLM-5.5 or use GLM-5.2 now? Use 5.2 now. It is the strongest value coder available ($1.40/$4.40, SWE-bench Pro 62.1), and an OpenAI-compatible setup makes adopting 5.5 at launch a one-line change. Waiting buys nothing.