What Happened
Markets & Semiconductors · Language Models · AI Capability · China · Market Reportreleased 26 Aug 2026 · record to 27 Aug 2026
A model that scores well, a claim nobody has verified, and a share price that answers neither
Z.ai's Chinese-Chip Claim for GLM-5.3-Flash
Z.ai (Zhipu AI) released GLM-5.3-Flash on Wednesday 26 August 2026 — a 320-billion-parameter mixture-of-experts model, previewed anonymously for a week under the code name “Ox Alpha.” Z.ai's own posts state the model “ran entirely on Chinese AI chips,” without naming a vendor, a chip model, or the “100,000” count the company has given only to reporters. CNBC states plainly that it “was unable to independently verify Z.ai's chip claims” and that Z.ai “declined to share details on which companies' chips it was using” — and separately notes that running a model takes far less computing power than training one. Z.ai's Hong Kong-listed shares closed more than 12% higher the same Thursday; rival MiniMax's own first-half results, filed the evening before, moved its shares too. What is independently confirmed, what is Z.ai's word alone, and what the market's reaction does and does not show.
- 57GLM-5.3-Flash's score on the Artificial Analysis Intelligence Index — #3 of 110 open-weights models in its size class, as of 27 Aug 2026
- 100,000chips Z.ai told reporters it used, on hardware it declined to name — not found in any of Z.ai's own dated public material
- 27.6Ttokens the model served on OpenRouter over 20–26 Aug 2026, combining its stealth-preview and named identities — more than double the second-placed model
- +12%Z.ai's own Thursday close, 27 Aug 2026 — CNBC's own report, filed earlier the same day, said “more than 8%”
- 283.1%MiniMax's own first-half 2026 revenue growth, year on year, disclosed the evening before Z.ai's own release
What's Confirmed, Graded each claim graded beside itself, not after it
| Standing | What | Detail |
|---|---|---|
| Confirmed | That Z.ai released GLM-5.3-Flash on 26 August, after a week-long anonymous preview | A 320-billion-parameter, 18-billion-active mixture-of-experts model (“320B-A18B”), natively multimodal, with a 1,048,576-token context window, released under the MIT licence with weights on Hugging Face; Z.ai's own pricing is $0.15 per million input tokens and $0.50 per million output tokens. It was tested anonymously for a week as “Ox Alpha” on OpenRouter and OpenCode, processing 62 trillion tokens before the formal release, Z.ai's own account says. |
| Confirmed, not verifiable here | That the model “ran entirely on Chinese AI chips” | Z.ai's own words, in its 26 August posts on X and on its blog; neither names a chip vendor, a chip model, or gives a count. CNBC's own report states plainly: “CNBC was unable to independently verify Z.ai's chip claims. The company declined to share details on which companies' chips it was using.” |
| Unconfirmed | The “100,000” chip figure, and that it covered all online requests including the Ox Alpha preview | Attributed to “the company” by CNBC and repeated by other outlets with no dated Z.ai statement behind it; not found in any of Z.ai's own published material. |
| Confirmed | That GLM-5.3-Flash outscores the models CNBC's report compares it with | GLM-5.3-Flash scores 57 on the Artificial Analysis Intelligence Index. The model CNBC's report calls “DeepSeek V4 Pro Max” scores 53 — “max” is DeepSeek's own reasoning-effort label for this variant, not a separate model name; Artificial Analysis's own page titles it “DeepSeek V4 Pro 0813 (max).” MiniMax M3 scores 45. 57 exceeds both. The site's full leaderboard lists every reasoning-effort variant of a model as its own row, which moves any single ordinal position depending on how they're counted; the report's “10th” and “18th” placements could not be exactly reproduced by that method — the scores themselves are exact. |
| Confirmed | That GLM-5.3-Flash led OpenRouter usage over the week | OpenRouter's own usage data tracks the model's anonymous-preview identity (“stealth/ox-alpha”) and its named release (“z-ai/glm-5.3-flash”) as separate slugs, because that is how the traffic was actually labelled at the time it ran. Combined across 20–26 Aug 2026, the two total about 27.6 trillion tokens — more than double the second-placed model — which is first place by usage for the week. A separate figure reported by the South China Morning Post, 10.3 trillion tokens “first among coding systems,” is a narrower claim about one usage category and is not the platform-wide total. |
| Confirmed | That Z.ai's shares closed sharply higher on Thursday | Z.ai's shares (2513.HK) closed at HK$1,160 on Thursday 27 Aug 2026, more than 12% higher on the day, per the South China Morning Post's own report, published after the market's 4pm close. CNBC's own report, filed earlier the same day, states “more than 8%” — an earlier, intraday figure from the same trading day, not a contradiction. Since Z.ai's own 8 Jan 2026 listing at an offer price of HK$116.20, that is a rise of about 898%. |
| Confirmed | That MiniMax's own first-half results carry two loss figures, moving in opposite directions | MiniMax's own reconciliation shows an adjusted net loss — with share-based payments, a fair-value loss and listing expenses added back — of $293.0 million, up 111.2% from $138.7 million a year earlier; this is the figure CNBC's report names. MiniMax's total, statutory net loss for the same six months narrowed 11%, to $358 million, per the South China Morning Post's own report of the same filing — a figure CNBC's report does not carry. |
Timeline
8 January to 27 August 2026, in Eight Moves everything with a date, in order
- 8 Jan 2026Z.ai lists on the Hong Kong Stock Exchange (2513.HK) at an offer price of HK$116.20.
- 9 Jan 2026MiniMax lists on the Hong Kong Stock Exchange (0100.HK) at an offer price of HK$165.00, closing its first day 109.1% higher.
- 20 Aug 2026Z.ai begins an anonymous preview of the model, under the code name “Ox Alpha,” on OpenRouter and the OpenCode agent platform.
- 26 Aug 2026Z.ai formally releases GLM-5.3-Flash, revealing it as the model previewed as Ox Alpha; its posts on X and its blog each state the model “ran entirely on Chinese AI chips,” without naming a vendor.
- 26 Aug 2026, later that dayMiniMax announces its first-half 2026 results: revenue up 283.1%, adjusted net loss up 111.2% to $293.0 million, total net loss narrowed 11% to $358 million.
- 27 Aug 2026Z.ai's shares close at HK$1,160, more than 12% higher on the day; CNBC's own report, filed earlier the same day, states “more than 8%.”
- 27 Aug 2026, the same dayMiniMax's shares rise in reaction to the previous evening's results — CNBC reports around 3%.
- 27 Aug 2026, also that dayCNBC publishes its report, stating it could not independently verify Z.ai's chip claim.
The Argument
What the Claim Can, and Cannot, Show the training-versus-inference distinction, stated plainly
Z.ai's claim rests on Z.ai's own words, and nowhere further. “Ran entirely on Chinese AI chips” appears in the company's 26 August posts on X and on its blog; the more specific “100,000 chips” figure that made headlines does not appear in either — it was given to reporters, and it does not appear in any dated material Z.ai has published. CNBC's own report is direct about the limit of what it could establish: “CNBC was unable to independently verify Z.ai's chip claims. The company declined to share details on which companies' chips it was using.” The same report states the fact that bounds what the claim would prove even if fully true: “Running an AI model requires less computing power than training a model.” Serving inference requests, at whatever scale, is a materially lower bar than training a frontier model on the same hardware — so “ran entirely on domestic chips” for inference does not, by itself, establish that China's chip ecosystem can support training at the same scale. Nor does the market's reaction stand in for verification. Z.ai's own shares closed more than 12% higher the same Thursday, but that is a fact about how investors responded to an announcement they could not check either — not independent evidence about which hardware served the requests.
- 0chip vendors or chip models named in any of Z.ai's own dated public material
- 57 : 53GLM-5.3-Flash's Intelligence Index score against the DeepSeek variant CNBC compares it with — confirmed independently of the chip claim
Named Readings, Not Findings an analyst's judgement, attributed and left as one
| Who | Says |
|---|---|
| Ivan Lam | Counterpoint Senior Research Analyst, quoted by CNBC: Z.ai is “likely using Huawei Ascend along with other chip suppliers,” and “more companies in China are working closely together for deeper collaboration across hardware and software” — his own reading, not independently confirmed by CNBC. |
| wccftech | Speculates separately, in its own reporting, that “Huawei's Ascend-class GPUs were likely powering” the run — its own inference, not sourced to Z.ai or to Lam. |
What Others Add
The Numbers Behind the Comparisons from the index and from OpenRouter's own data
Intelligence Index
Ahead of the models CNBC names, though not by the ranks CNBC gives
- GLM-5.3-Flash scores 57 on the Artificial Analysis Intelligence Index — #3 of 110 within its own size-and-openness class (open-weights, over 150 billion parameters), as of 27 Aug 2026.
- DeepSeek V4 Pro 0813 (“Max Effort”) — the model CNBC's report calls “DeepSeek V4 Pro Max” — scores 53. “Max” is DeepSeek's own reasoning-effort label for this variant, not a separate model name.
- MiniMax M3 scores 45 on the same index. 57 exceeds both 53 and 45; the comparative claim CNBC's report makes holds, on the scores themselves.
- The site's full leaderboard lists every reasoning-effort variant of a model as its own row, which moves any single ordinal position depending on how they're counted; the report's “10th” and “18th” placements could not be exactly reproduced by that method.
OpenRouter Usage
First by usage, once the stealth-preview identity is added in
- OpenRouter tracks the model's anonymous-preview identity (“stealth/ox-alpha”) and its named release (“z-ai/glm-5.3-flash”) as separate slugs, because that is how the traffic was actually labelled at the time it ran.
- Combined across 20–26 Aug 2026, the two total about 27.6 trillion tokens — more than double the second-placed model — confirming first place by usage for the week.
- A separate figure from the South China Morning Post — 10.3 trillion tokens, “first among coding systems” — is a narrower claim about one usage category, not the platform-wide total.
MiniMax's Own First-Half Results two loss figures, moving in opposite directions
| What | MiniMax's Own Figure |
|---|---|
| Revenue, six months to 30 June 2026 | $116.6 million, against $30.4 million a year earlier — up 283.1%, already exceeding MiniMax's entire FY2025 revenue of $79.0 million. |
| Enterprise/Open Platform revenue | Rose 703.1%, from $9.2 million to $73.9 million, and from 30.3% to 63.4% of total revenue — the segment MiniMax itself credits with driving the headline growth. |
| Gross profit | Rose 464.8%, from $3.7 million to $20.8 million; margin rose from 12.1% to 17.9%. |
| Adjusted net loss | Up 111.2%, from $138.7 million to $293.0 million — adds back share-based payments, a fair-value loss and listing expenses. This is the figure CNBC's report names. |
| Total (statutory) net loss | Narrowed 11%, to $358 million, over the same six months — the opposite direction from the adjusted figure, and a figure CNBC's report does not carry. |
| Cash balance | $1,322.8 million at 30 June 2026, against $1,050.3 million at 31 December 2025. |
Conclusion
Where This Stands what's settled, and what depends on a vendor nobody has named
What the record supports
That Z.ai released GLM-5.3-Flash on 26 August, after a week-long anonymous preview as “Ox Alpha” that itself carried more than 27 trillion tokens of OpenRouter traffic. That the model scores 57 on the Artificial Analysis Intelligence Index, ahead of the DeepSeek and MiniMax models CNBC's report names, though not at the exact ranks the report gives. That Z.ai's shares closed more than 12% higher, and MiniMax's own first-half results — filed the evening before, and carrying two loss figures moving in opposite directions — moved MiniMax's shares the same week.
What it doesn't yet
Which company's chips Z.ai actually used — not Z.ai (declined to say), not CNBC (could not verify), and no vendor, model number or independent count has appeared anywhere. The exact ordinal placements CNBC's report gives on the Intelligence Index. Z.ai's own first-half 2026 results, due at a board meeting scheduled for 31 August 2026 and not yet published as of 27 August 2026. And, regardless of how any of that resolves, running a model — at whatever scale, on whatever hardware — is not the same test as training one.
What Would Settle It the concrete tells
- Z.ai naming a chip vendor and a chip model, rather than the general “Chinese AI chips” its own posts use.
- An independently sourced count behind the “100,000” figure — from Z.ai's own dated material, not only from what the company told reporters.
- Z.ai's outstanding first-half 2026 results, due at a board meeting scheduled for 31 August 2026.
- A published Intelligence Index ordinal position, once repeated reasoning-effort rows for the same model are collapsed to one, that matches or corrects the report's “10th.”