Three Chinese frontier releases in one week: Kimi K3, Qwen 3.8-Max, DeepSeek V4
Three frontier-class Chinese releases in seven days. Open weights at 2.8T, a Fable-5-adjacent claim from Alibaba, and DeepSeek V4 going GA. Below: what shipped, and what actually changes for production teams.
Moonshot Kimi K3: 2.8T open-weight, ranked #3 in intelligence globally
Kimi K3 is 2.8 trillion parameters, open weights, 1M-token context. Artificial Analysis ranks it third overall in its intelligence index, behind only GPT-5.6 Sol and Fable 5. First time a fully open Chinese model has cracked the global top three. Marks Moonshot's pivot away from cost-optimized models toward true frontier capability.
Open weights at frontier quality means every US enterprise can now run top-3 inference on its own hardware. No API dependency, no rate limits, no export-control exposure.
Read at Bloomberg → link
Alibaba Qwen 3.8-Max preview: 2.4T multimodal, "second only to Fable 5"
Announced July 19. 2.4 trillion parameters, natively multimodal across text, image, video, and documents. Alibaba's positioning: "one of the most powerful models available today, comparable to leading frontier AI models, second only to Fable 5." Available via Alibaba's Token Plan, Qoder, and QoderWork. Hong Kong shares climbed as much as 5% on the announcement.
The Fable 5 comparison is Alibaba's, not independent — but it's the first time a Chinese lab has publicly framed itself as adjacent to the current SOTA. That framing itself is the news.
Read at MarkTechPost → link
DeepSeek V4 goes GA: 1.6T Pro and 284B Flash, both 1M context
The April V4 preview is graduating to final release this month. Two variants: V4 Pro at 1.6T total with 49B active per token (MoE), and V4 Flash at 284B total with 13B active. Both ship with 1M-token context, both open source. Biggest DeepSeek release since R1 in January 2025, following the sparsity-first architecture line that lets V4 run on Ascend at meaningfully lower FLOPs than V3.
V4 is the open-weights option that pairs a frontier model with genuine Chinese-silicon portability. For teams evaluating US chip risk, this is the reference stack.
Read at ExplainX → link
US enterprises now standardize on Qwen, GLM, and Kimi at production scale
American tech companies have shifted to Chinese model families — Alibaba's Qwen, Z.ai's GLM, Moonshot's Kimi — at what WaPo calls "landmark adoption" scale. A year ago, using a Chinese model in a US corporate stack was unthinkable. The pull is boring and durable: GLM-5.2 delivers near-GPT-5 quality at a fraction of the price, Kimi K3's open weights beat any closed US model on deployability, and Qwen plus GLM have quietly become dev favorites for coding.
If you are pricing an FY26 inference budget, the "Chinese model in the stack" line item is now a mainstream option, not a curiosity.
Read at Washington Post → link
Huawei Atlas 950 SuperPoD debuts as domestic training hardware
Huawei publicly showcased the Atlas 950 SuperPoD, its new AI computing system, in the same week as Kimi K3. Not yet at Blackwell parity, but the signal is that China increasingly has the domestic compute stack it needs to keep training frontier-class models despite US chip export restrictions. Consistent with the V4-on-Ascend story: chip-model coupling is the throughline.
The "we don't need Nvidia" narrative is starting to hold up. Watch for whether US export policy responds.
Read at KSAT / AP → link