Amazon Bedrock Adds Chinese Models Kimi and GLM; Jassy Says West 'Very Hungry' for Western Open Models

Amazon CEO Andy Jassy announced during the company's third-quarter earnings call on Thursday that AWS Bedrock will host two Chinese frontier models—Moonshot AI's Kimi K3 and Z.ai's GLM 5.3. The additions were disclosed alongside Bedrock's existing lineup of closed models from Anthropic and OpenAI. Jassy also referenced what the Benzinga report called "SpaceXAI's Grok 4.7" — likely a reference to xAI's Grok model, as SpaceX does not develop AI models.

What was announced

According to a report on the earnings call by Benzinga, Bedrock now offers Moonshot's Kimi K3 and Z.ai's GLM 5.3 as open-model options. Jassy described Bedrock's potential as "gigantic," suggesting it could eventually reach a scale comparable to EC2, AWS's core compute service. He also said the platform's security architecture ensures that customer inputs and outputs are not shared with model providers or used to train their foundational models. Jassy predicted that within two to three years, six to eight models will handle more than 95 percent of workloads on Bedrock. [4]

The earnings call referenced "Kimi K3" and "Z.ai's GLM 5.3," while separate reporting about model training cites "Kimi K3.1 Pro" and Zhipu AI's GLM. It is unclear whether the Bedrock offering is the base Kimi K3 or the updated Kimi K3.1 Pro referenced in distillation reports. The sources use different names — Z.ai on the earnings call, Zhipu AI in other coverage — without explaining the relationship. [1][4]

Jassy on open-model demand

"The Western world is very hungry for additional open models that are Western open models," Jassy said, adding, "I believe that will happen over time as well." The phrasing — "Western open models" — is notable: it frames the demand as being for open-weight models developed in the West, not simply for open weights regardless of origin. [4]

Jassy's comment about Western demand for open models comes as Chinese open-weight systems show competitive performance on benchmarks.

Chinese model performance context

In October, Artificial Analysis ranked Mistral Large 4 as the most powerful non-Chinese open-weight model, but noted it "edges out" the Chinese GLM-5.3 only narrowly — and uses 47 percent fewer tokens to do so. [1]

The Decoder reported that Anthropic said GLM-5.3 writes exploits nearly as well as Claude Mythos Preview, its restricted-access offensive-cybersecurity model. The Flash variant of GLM-5.3 assembled a reliable Chrome exploit for $20.40 at Zhipu's API prices, and CAISI, the U.S. government evaluation agency, confirmed the result. [3]

Epoch AI separately demonstrated GLM-5.3 autonomously writing functional exploits for six recent Firefox vulnerabilities using only publicly available patch data, with the fastest run completing in 12 hours. [1]

Broader strategic signals

Former Google CEO Eric Schmidt said earlier in October that China trails the U.S. by roughly six months in AI but warned that free, open-weight Chinese models could gain strategic traction in developing markets. He attributed part of China's progress to access to restricted advanced chips and the use of distillation — training on outputs of Western models — to replicate capabilities. [4]

Nvidia CEO Jensen Huang called Chinese models "excellent" in July and argued Washington should use rather than ban high-quality open-source technology, saying wider adoption of models like DeepSeek and Kimi would boost demand for Nvidia chips. [4]

What remains unclear

• First-of-kind claim: Available reporting does not independently verify that this is the first time a major U.S. cloud provider has offered Chinese frontier models as managed services.
• Model versions: The earnings call cited "Kimi K3" and "GLM 5.3"; other coverage refers to "Kimi K3.1 Pro" and Zhipu AI's GLM. It is not clear whether these refer to the same releases or incremental updates.
• Security architecture: Jassy's assertion that customer data is not shared with model providers or used for training is a claim about Bedrock's design; no independent audit or technical detail was provided in the sources.
• Consolidation forecast: The prediction that six to eight models will cover 95 percent of workloads within two to three years is a forward-looking statement without disclosed methodology.
• Distillation allegations: The Information reported that DeepSeek and Moonshot used outputs from OpenAI's GPT-6 Astra to train recent models including Kimi K3.1 Pro, a practice both OpenAI and Anthropic prohibit. Moonshot did not comment; DeepSeek did not respond. [1]

Financial backdrop

Amazon reported Q3 net sales of $201.9 billion and AWS revenue of $46.0 billion, up 39 percent year-over-year. Trailing twelve-month free cash flow was a $42.4 billion outflow, driven by a 2026 capital expenditure outlook of approximately $220 billion. [2]

The Bedrock expansion comes as Amazon deepens its infrastructure bet: Anthropic has committed to spending $100–110 billion on AWS over ten years (sources differ on the exact figure) as part of Amazon's investment in Anthropic, which itself has filed publicly for an IPO. [2][3]