- Bloomberg Intelligence measured the performance gap between Chinese and US AI models at 6% in June, a record low, down from 9% in May.
- Moonshot’s Kimi K3 — a 2.8-trillion-parameter open-weights model released July 17 — followed Z.ai’s GLM-5.2 debut in June, which BI says makes the trend structural rather than a single outlier.
- The K3 release sent Hong Kong-listed Chinese AI peers down sharply: Zhipu fell 27.7% and Minimax 16.5%.
- Bloomberg Intelligence also expects Chinese firms to shift 46% of AI accelerator budgets to domestic suppliers over the next 12 months, up from 30%.
What Happened
Chinese AI models narrowed the measured performance gap with US counterparts to 6% in June, down from 9% in May and the smallest margin recorded, according to Bloomberg Intelligence analysis reported by Bloomberg on July 23, 2026. BI’s framing is that Moonshot’s Kimi K3 confirms Zhipu’s earlier showing was not an isolated result, which it says raises questions about the sustainability of US technological supremacy in the sector.
The two data points behind the reading are Z.ai’s GLM-5.2, which debuted in June scoring near top US closed-source models on benchmark tests, and Moonshot’s Kimi K3, released July 17 as a 2.8-trillion-parameter open-weights model.
Why It Matters
The prevailing estimate through early 2026 was that Chinese labs trailed US frontier systems by at least six months. As recently as last week an Anthropic executive put Claude’s lead at roughly six to twelve months — an assessment that the K3 release immediately put under pressure.
The competitive channel is weights, not just scores. Kimi K3 and GLM-5.2 ship open, so a narrowing benchmark gap translates directly into deployable capability for anyone willing to self-host, which is a different threat model than a closed model matching US performance behind an API. That is also why the release moved markets: Moonshot’s own Chinese peers took the damage, with Zhipu down 27.7% and Minimax down 16.5% in Hong Kong.
Technical Details
The 6% figure is a composite performance gap computed by Bloomberg Intelligence, not a single benchmark score, and it moved three percentage points in one month — a magnitude that reflects the lumpy nature of frontier releases rather than steady convergence. A single flagship launch can compress it, and a US release can widen it again.
Kimi K3’s 2.8 trillion parameters make it the largest open model released to date. Parameter count is not a performance measure on its own, but at that scale the relevant constraint is training compute, which is precisely where US export controls were meant to bind — and which is why the model’s provenance became a policy question within days.
Bloomberg Intelligence separately projects that Chinese firms will allocate 46% of AI accelerator budgets to domestic suppliers over the next 12 months, up from 30% currently, indicating the substitution away from restricted Nvidia parts is already priced into procurement plans.
Who’s Affected
US frontier labs face the pricing consequence first: an open model within a few percent of closed-model performance sets a ceiling on what mid-tier API capability can be charged for. Enterprises evaluating self-hosted deployments gain a credible option they did not have two quarters ago.
Chinese vendors are affected unevenly, and the market reaction showed it. A single lab pulling ahead compresses the field behind it rather than lifting it, which is why Zhipu and Minimax sold off on a competitor’s release. The 46% domestic-accelerator projection points the same way: procurement is consolidating around whichever domestic stack can actually train at frontier scale.
Policymakers face the harder question. On July 22, White House OSTP Director Michael Kratsios alleged Moonshot had trained K3 partly through large-scale distillation of Anthropic’s Fable model and had accessed restricted Nvidia GB300 hardware — allegations that, if substantiated, would mean the gap narrowed through routes the export-control regime was designed to close.
What’s Next
BI’s measure is monthly, so the July reading — which will capture K3 in full and any US response — is the next observation. The gap is narrow enough now that a single frontier release on either side will move it materially, which argues against reading any one month as a trend on its own.
