China AI

AIDeepSeek has open-sourced DeepInfer under Apache 2.0, the inference engine it says runs its production services, and released deepseek-recipe under MIT. The framework supports only DeepSeek-V4.1-Flash on NVIDIA SM100 and Huawei Ascend 950 and lacks a KV cache system, P/D disaggregation, and online serving.

Why it matters: The article compares DeepInfer's features and limits with vLLM, SGLang, and TensorRT-LLM, which helps readers judge its fit for their own serving stack.

AITenstorrent, an inference-focused chip startup, has raised $1.5 billion at a $6.2 billion valuation, more than double its late 2024 valuation.

Why it matters: The round's valuation history and the rejected acquisition interest show how investors now value AI inference chip startups competing with Nvidia.

AIThe Department of Homeland Security has ordered US civilian agencies to remove DeepSeek and DeepSeek-affiliate products, plus third-party products built on DeepSeek implementations, under the Federal Acquisition Supply Chain Security Act of 2018.

Why it matters: The source ties the order to a federal notice and its legal basis, showing how supply-chain security rules can reach AI models even when they are hosted in the US.

AIArtificial Analysis says DeepSeek V4.1 Flash 1017 scores 45 on its Intelligence Index, six points above DeepSeek V4.1 Flash, and ties GLM-5.3 as the #2 open-weights model.

Why it matters: The post places the update against peer models on cost and capability, showing where it sits among open-weights options and what it costs per task.

AIMiniMax released M3.1 Flash, which it says reaches leading-model levels on agentic, coding, computer use, and cybersecurity benchmarks.

Why it matters: The post lists benchmark scores against named competitors and gives per-million-token prices, so readers can compare cost and capability directly.

AIAlibaba's Tongyi MAI team has open-sourced OpenGCUA, a full stack for a generalist computer-use agent, including a harness, cross-device SDK, native macOS app, evaluation trajectories, and a reference agent.

Why it matters: The post explains how OpenGCUA lets models choose observation, action, and task organization, with OSWorld-v2 results and a real-device token-cost caveat.

AIvLLM reports Kimi K3 decode reaching 842 tokens/s at batch 1 on NVIDIA GB300 NVL72 using DSpark.

Why it matters: The post gives concrete throughput, interactivity and configuration results, showing how a unified decode kernel changes serving performance for a large model.

AIDeepSeek releases DeepSeek-V4.1-Flash-1017 on Hugging Face as a minor update to DeepSeek-V4.1-Flash.

Why it matters: The source gives benchmark scores against several named rival models, showing where this release sits on performance and cost in agent tasks.

AIvLLM reports a new Kimi K3 decode path built with agent-assisted CuTe DSL kernels for NVIDIA GB300 GPUs.

Why it matters: The post explains how agent-assisted kernel work is paired with isolation, validation, and reproducible benchmarks so optimized inference code can be reviewed and maintained.

AIChina's commerce and science ministries released a draft update to its export control catalog covering 23 AI technologies.

Why it matters: The article links China's draft export rules to concrete thresholds and exemptions, and explains how they could affect open-weight model access in the US and Europe.

AIArena says the largest gap between AI labs' agents is whether users enjoy working with them, not whether the agents complete tasks.

Why it matters: The post separates user satisfaction from task success across agents, showing where labs differ and where they are converging.

AITencent Hunyuan releases Hy-Memory-30B-A3B, a compact writer model with about 3B active parameters that decides what to retain in external memory frameworks such as Mem0, MemOS, and Hy-Memory.

Why it matters: The model card lays out a two-stage training method and cross-harness benchmark results, making it useful for comparing memory writers against general-purpose LLMs.

AIKuaishou's Kling AI says Kling 4.0 is now fully available on klingai.com, with native clips up to 30 seconds and audio included. The company says 4K and 1080p 10-bit HDR output and an enterprise API are coming soon.

Why it matters: The article separates features new to the public rollout from those already previewed in September, showing which capabilities are available now and which remain pending.

AIAMD and Embedded LLM report that an eight-GPU MI355X system serving Kimi K3 with vLLM delivers 24.4% more total tokens per second per GPU than NVIDIA HGX B300 at 50 tokens/s/user P90 interactivity on the AgentX benchmark.

Why it matters: The post traces optimizations for serving Kimi K3 on AMD MI355X, linking vLLM serving choices to GPU kernels and hardware limits, useful for diagnosing bottlenecks in other inference stacks.

AIStepFun has released the BF16 and FP8 weights of Step 5 Preview on Hugging Face and ModelScope, according to RadixArk's post.

Why it matters: The post identifies the open-weight release channels and the serving stack, showing how a 600B-parameter MoE model is deployed through SGLang in production.

AIMeta releases Muse Spark Open, an open-weight model with 816 billion parameters (42 billion active) and up to one million tokens of context.

Why it matters: The report gives benchmark scores against Qwen, GLM, and Kimi, along with hardware and license terms, so readers can judge where the model stands in the open-weight race.

AISGLang announces Day-0 support for StepFun's Step 5 Preview, a 600B-parameter sparse MoE model with 27B active per token and a 1M-token context window.

Why it matters: The post names the inference and training setup, including hybrid attention and Day-0 RL support, showing how the open-weight model is served and trained.

AIApple executive chairman Tim Cook said in Shanghai that Apple Intelligence with Alibaba's Qwen model has been approved and a beta will launch in China later in October.

Why it matters: The report ties Apple Intelligence's China rollout to an approved Alibaba Qwen model, clarifying which features launch first and which are excluded.

AIShanghai AI Laboratory and Professor Xu Disheng's team say the Shusheng-S2 scientific multimodal model, working with human researchers, produced a proof of the three-dimensional Ghys conjecture.

Why it matters: The article details how humans and the Shusheng-S2 model worked iteratively on a three-dimensional Ghys conjecture result, and how Peking University verified it in Lean.

AISkywork announces Matrix-Game 4.0, an open-source video model that lets players interact with a generated world.

Why it matters: The post gives concrete speed and hardware figures, and links the code, weights and demos, so readers can check whether a playable video world runs on their own GPU.