Model scale and architecture: Kimi K3.1 Pro has 2.8 trillion total parameters, approximately 104 billion active parameters, and a one-million-token context window. 4 Reporting on Alibaba’s announcement describes sparse MoE, hybrid KDA linear attention and DSA sparse attention, and multi-token prediction. 9

Joint adaptation: The announcement reporting says Alibaba Cloud, T-Head, and the Kimi team jointly adapted the model using the T-Head SAIL software stack; the available excerpt also explicitly mentions FlashMLA optimization. 9 Insufficient evidence in the retrieved excerpts to confirm the additional DeepGEMM MoE, MegaMoE, and MoonEP adaptation details.

Independent framework confirmation: SGLang’s Kimi K3.1 Pro cookbook explicitly lists PPU support for M890 and 810E, independently corroborating hardware support beyond Alibaba’s own announcement. 4 It identifies a kimi-k3.1-day-0 branch based on main, so this evidence should not be presented as proof that all launch-day changes were already merged into the main branch. 4

Deployment and location: Alibaba’s announcement confirms deployment on its SuperPoD instances and invited testing in Zhongwei; contemporaneous reporting also says the model was deployed on the Qwen AI platform with an API service. 159 Insufficient evidence to verify bare-metal instance specifics, REIN service deployment, or day-zero availability in Ulanqab.

Next-generation processor: Alibaba’s separate Apsara announcement claims the Zhenwu V900 delivers three times M890’s performance, with 216 GB of memory and 1,200 GB/s inter-chip bandwidth. 16 This is a vendor performance claim, not an independently verified Kimi K3.1 Pro benchmark.

Memory and interconnect: As an engineering inference, sparse activation reduces computation per token but does not eliminate the need to store the model’s weights; distributing a model of this scale also makes expert-routing communication and long-context state management important. The retrieved evidence provides no measured memory footprint, required accelerator count, or scale-up bandwidth threshold.

“First day-zero” and serving recipes: Insufficient evidence to establish that this was Zhenwu’s first mainstream open-weight day-zero launch through an in-tree hardware backend, or to reproduce verified SGLang prefill/decode recipes with multi-node tensor and expert parallelism. The cookbook excerpt establishes hardware coverage and the day-zero branch, not those stronger claims. 4

Panmai 920 and adoption: I don’t know from the available excerpts whether the announcement documented PCIe 5.0, 112G PAM4, 400Gbps throughput, 400 million packets per second, multipath RDMA, programmable congestion control, an integrated GPU/SSD PCIe switch, or roughly 30% system-cost savings. The “more than 100 large customers” claim and its industry breakdown are likewise unverified here.

Huawei comparison: A November 4 SGLang issue describes Kimi K3.1 support work on Ascend, including routing its DSA attention path. 14 That demonstrates a separate Ascend integration effort, but is insufficient evidence of a completed Huawei SuperPoD launch-day offering or a second production-ready domestic stack.

Model scale and architecture: Kimi K3.1 Pro has 2.8 trillion total parameters, approximately 104 billion active parameters, and a one-million-token context window. 4 Reporting on Alibaba’s announcement describes sparse MoE, hybrid KDA linear attention and DSA sparse attention, and multi-token prediction. 9

Joint adaptation: The announcement reporting says Alibaba Cloud, T-Head, and the Kimi team jointly adapted the model using the T-Head SAIL software stack; the available excerpt also explicitly mentions FlashMLA optimization. 9 Insufficient evidence in the retrieved excerpts to confirm the additional DeepGEMM MoE, MegaMoE, and MoonEP adaptation details.

Independent framework confirmation: SGLang’s Kimi K3.1 Pro cookbook explicitly lists PPU support for M890 and 810E, independently corroborating hardware support beyond Alibaba’s own announcement. 4 It identifies a kimi-k3.1-day-0 branch based on main, so this evidence should not be presented as proof that all launch-day changes were already merged into the main branch. 4

Deployment and location: Alibaba’s announcement confirms deployment on its SuperPoD instances and invited testing in Zhongwei; contemporaneous reporting also says the model was deployed on the Qwen AI platform with an API service. 159 Insufficient evidence to verify bare-metal instance specifics, REIN service deployment, or day-zero availability in Ulanqab.

Next-generation processor: Alibaba’s separate Apsara announcement claims the Zhenwu V900 delivers three times M890’s performance, with 216 GB of memory and 1,200 GB/s inter-chip bandwidth. 16 This is a vendor performance claim, not an independently verified Kimi K3.1 Pro benchmark.

Memory and interconnect: As an engineering inference, sparse activation reduces computation per token but does not eliminate the need to store the model’s weights; distributing a model of this scale also makes expert-routing communication and long-context state management important. The retrieved evidence provides no measured memory footprint, required accelerator count, or scale-up bandwidth threshold.

“First day-zero” and serving recipes: Insufficient evidence to establish that this was Zhenwu’s first mainstream open-weight day-zero launch through an in-tree hardware backend, or to reproduce verified SGLang prefill/decode recipes with multi-node tensor and expert parallelism. The cookbook excerpt establishes hardware coverage and the day-zero branch, not those stronger claims. 4

Panmai 920 and adoption: I don’t know from the available excerpts whether the announcement documented PCIe 5.0, 112G PAM4, 400Gbps throughput, 400 million packets per second, multipath RDMA, programmable congestion control, an integrated GPU/SSD PCIe switch, or roughly 30% system-cost savings. The “more than 100 large customers” claim and its industry breakdown are likewise unverified here.

Huawei comparison: A November 4 SGLang issue describes Kimi K3.1 support work on Ascend, including routing its DSA attention path. 14 That demonstrates a separate Ascend integration effort, but is insufficient evidence of a completed Huawei SuperPoD launch-day offering or a second production-ready domestic stack.