Stealth Model Watch
Confirmation · Aug 26, 2026
Ox Alpha Revealed: Z.AI Confirms a New GLM-Series Model
On August 26, 2026 — Day 7 of the stealth preview — Z.AI Co. (Zhipu) told Bloomberg News that Ox Alpha is a new iteration of its GLM series. On the night of Aug 26, Z.AI formally announced and open-sourced it as GLM-5.3-Flash (320B-A18B) — the anonymous preview’s official name. The Chinese community knew it as “牛来”.
- GLM series — officially confirmed
- Open weights — released Aug 26
- $0 preview — still free
What we know
The official confirmation, in three layers
Ranked by credibility: an on-record statement to a newsroom, a named executive's product claim, and what does not exist yet.
Layer 1 · Z.AI confirms, on the record
Verified factZ.AI confirmed to Bloomberg News — in response to a query, not through its own channels — that Ox Alpha is “a new iteration of its GLM series” and would “release the weights for it tonight”. What that confirmation settles:
- Identity: a new iteration of Z.AI’s GLM series.
- Open weights: released the same night as GLM-5.3-Flash (Aug 26).
- Price: the model stays free during the preview.
- Traction: #1 on OpenRouter’s leaderboard — roughly double DeepSeek’s usage (23.2T vs 11.6T tokens at the rankings snapshot).
- Profile: a reasoning model built for coding and agentic work, with text, image and video input.
Source: Bloomberg News, Aug 26, 2026 (reporter Luz Ding).
AS REPORTED
China’s Z.AI made Ox Alpha stealth model that rivals DeepSeek
Bloomberg News · Luz Ding · Aug 26, 2026 — via The Edge Malaysia
Image: The Edge Malaysia’s publication of the Bloomberg report (Aug 26, 2026).
Layer 2 · The official model: GLM-5.3-Flash
Verified factZ.AI announced the model as GLM-5.3-Flash (320B total, 18B active) — calling it “the first native multimodal model in the GLM-5 series.” The name was first reported by Nebius co-founder Roman Chernin on X; Z.AI made it official on Aug 26.
- AA Intelligence Index 57 — level with Claude Opus 4.8.
- Pricing: 1/10 of GLM-5.3 (1/20 in the launch window, 1/40 of Opus 4.8).
- Compute: the entire preview ran on Z.AI’s domestic chip cluster.
ON THE PARETO FRONTIER
GLM-5.3-Flash: 57 points at $0.045 per task — best price-performance on the index
Artificial Analysis Intelligence Index v4.1.1 · updated Aug 26, 2026
Layer 3 · The open weights
Verified factThe weights shipped to Z.AI’s official Hugging Face org on Aug 26, 2026: zai-org/GLM-5.3-Flash ↗ — released under the MIT license. Model card and config live on the repo page.
How we got here
The seven-day hunt
From anonymous launch to named developer in one week — reconstructed from the live evidence log.
-
Day 1 · Aug 20
Anonymous, free, and everywhere at once
Ox Alpha appeared on OpenRouter under the anonymous id stealth/ox-alpha — no lab, no announcement, no attribution. The same day, OpenCode listed it free for a week at near-unlimited limits. The hunt started with one question: who serves a frontier model and refuses to sign it?
-
Day 1–2 · Fingerprinting
Every hard probe converges on GLM
Adversarial probe strings matched the published GLM tokenizer 13 out of 13; a 25-prompt sweep tracked GLM-5.3 within a constant ~75-token wrapper offset; a reasoning_effort rejection returned the [1210] error code Z.ai’s API uses verbatim; even a malformed-image failure answered in Chinese.
-
Day 2 · Counter-evidence
The race stays open
Not every test agreed: some edge-case glitches looked MiMo-like instead of GLM, and commenters argued only a hyperscaler could give a model away at this scale. A viral 10-task eval putting Ox Alpha at 80% spread fast — useful early signal, tiny sample.
-
Day 2–3 · Measurement
The 80% rumor dies
An independent developer ran every one of the 113 DeepSWE tasks in a single 20-hour sweep: 58.4%, level with Claude Opus 4.8. The rumored ~80% pass rate was dead — but the measured profile, elite agentic coding on a million-token context, remained.
-
Day 4 · Mainstream coverage
Attention, still no attribution
TechCrunch and Bloomberg both wrote up the mystery model on Aug 23 — neither could put a developer behind it, and OpenRouter still described the provider as anonymous during the preview.
-
Day 7 · The reveal
Z.AI steps forward
Z.AI told Bloomberg News that Ox Alpha is its GLM series, promised open weights that night, and confirmed the preview stays free. Seven days after launch, the identity question was closed.
The community’s evidence chain had pointed at GLM from hour one — the official confirmation made circumstantial proof verified fact.
WATCH
The full test that called it: “a new GLM model?”
An independent hands-on review published during the anonymous window — its conclusion aged well. Plays here without leaving the page.
Stealth Ox Alpha: a new GLM model? Full test (YouTube) — plays here, no need to leave.
Model access
Weights & model access
Official weights were released on the night of August 26, 2026 under the name GLM-5.3-Flash (320B-A18B), in Z.AI’s official Hugging Face org (zai-org). It is the first native multimodal release in the GLM-5 series.
Hosted endpoints remain the quickest way to try it: OpenRouter
(stealth/ox-alpha) and OpenCode Zen
(x-preview-f-free), both at $0 during the
preview. Post-preview pricing is set at 1/10 of GLM-5.3 (1/20 in the launch window).
Weights live
GLM-5.3-Flash — open weights released
MIT license · 320B-A18B · FP8 checkpoint (BF16 sibling)
huggingface.co/zai-org/GLM-5.3-Flash ↗What the official release confirms
320B total · 18B active
GLM-5.3-Flash (320B-A18B). Sparse + linear attention hybrid, mHC connections, 30T multimodal pretraining tokens.
FP8 checkpoint + BF16
Official FP8 checkpoint is the main repo, with a BF16 sibling at zai-org/GLM-5.3-Flash-BF16. Community GGUF quants not posted yet.
VRAM
Z.AI has not published a runtime VRAM figure. The FP8 repo is ~656 GB on disk; serving VRAM depends on the framework.
The community’s first questions after any open-weight drop are size, VRAM and quants — this page answers them the moment the repo lands.
Verified fact
The backdrop: Stripe buys OpenRouter
The stage Ox Alpha topped was itself changing hands mid-preview.
Stripe’s acquisition of OpenRouter (announced Aug 19, reported at $7.5B–$8B) had just closed when Ox Alpha launched — the biggest launch in OpenRouter’s history landed weeks into the new ownership. Stripe CEO Patrick Collison called the stealth release “very impressive.”
After the reveal
What’s next
Three things left to watch now that the name is public.
-
Open weights — landed ✓
GLM-5.3-Flash shipped to Hugging Face on Aug 26 (zai-org/GLM-5.3-Flash, MIT license, FP8 + BF16). Community GGUF quants are next to watch.
-
Post-preview pricing Official
Pricing is now official: 1/10 of GLM-5.3 (1/20 in the launch window, 1/40 of Claude Opus 4.8).
-
Real-name listings across labs and aggregators
Plan rosters adding the model under its real name instead of an anonymous slot.
This page is the permanent archive of the Ox Alpha reveal — GLM-5.3-Flash is now public and open-weight.