Selected industry news from AIHOT. Follow advances in AI, products and technology.
AIHOT · Selected summaries with original sources · Nov 4, 2026, 9:00 AM
Moonshot AI open-sourced the Kimi Vendor Verifier (KVV) project alongside the release of the Kimi K2.6 model to verify the accuracy of open-source model inference implementations; it includes Pre-Verification.
PyTorch 2.10.0 has been officially released. Major updates include experimental support for Python 3.14 (including freethreaded builds) and full compatibility of torch.compile() with Python 3.14.
QwenTeam has open-sourced the full series of Qwen3-TTS voice models, including two sizes (1.7B and 0.6B). It supports voice creation, voice cloning, multilingual speech generation (covering 10 languages including Chinese, English, Japanese, and Korean), and acoustic attribute control based on natural language instructions.
During pre-production testing, Mistral AI discovered a memory leak growing at 400 MB per minute when using vLLM for Prefill/Decode disaggregated serving with the Mistral Medium 3.1 model and NIXL.
ElevenLabs launched the music album 'The Eleven Album', featuring Liza Minnelli, Art Garfunkel, KondZilla, IAMSU!
IBM Research与Hugging Face联合发布AssetOpsBench，一个专为工业资产生命周期管理设计的AI智能体评测基准，涵盖230万传感器数据点。
The official Hugging Face blog published the first article in a series reviewing the impact of the DeepSeek R1 model's release on the global open-source AI ecosystem one year later; it notes that R1 lowered technical barriers (by opening up reasoning paths).
vLLM has released version v0.14.0. Core changes include async scheduling being enabled by default (can be disabled), the addition of a gRPC server entry point, and `--max-model-len auto` for automatic GPU memory adaptation.
Overworld has released Waypoint-1, an interactive video diffusion model that supports real-time control via text, mouse, and keyboard, based on a frame-causal rectified flow transformer architecture.
Zhipu has officially open-sourced and made available for free the GLM-4.7-Flash hybrid thinking model, with a total of 30B parameters and 3B active parameters, achieving state-of-the-art (SOTA) scores among open-source models on benchmarks such as SWE-bench Verified and τ²-Bench.
DeepMind has released D4RT (Dynamic 4D Reconstruction and Tracking), a unified encoder-decoder model based on Transformer.
ChatGPT Go has launched globally, offering GPT-5.2 Instant, higher usage limits, and extended memory capabilities.
Google Research published the paper "Smartwatch-Based Walking Metrics Estimation," proposing a multi-output deep learning model based on Temporal Convolutional Networks (TCN).
MasterClass launched its On Call service, providing realistic voices for AI instructors such as Gordon Ramsay, Mark Cuban, and Chris Voss via ElevenLabs Text-to-Speech technology, supporting real-time voice conversations; 75% of users chose voice interaction, with significantly higher retention and engagement among voice users; the service includes two main modules—AI coaching guidance and AI role-playing—and employs a multi-layered safety framework to ensure brand consistency and content security.
Hugging Face and OpenAI have jointly launched the open reasoning standard "Open Responses," based on the Responses API released in March 2025. It extends support for native tool calling, sub-agent loops, structured output, and encrypted reasoning streams for Agent scenarios. The release includes curl examples, comparisons of event stream formats, and explanations of routing mechanisms, with early access available on Hugging Face Spaces.
Google released the MedGemma 1.5 4B multimodal large language model, which for the first time supports input of 3D CT, MRI, and whole-slide pathology images, improving disease classification accuracy by 3%–14% over MedGemma 1 on internal benchmarks.
Google DeepMind has released Veo 3.1 Ingredients to Video, which supports video generation based on reference images. It adds native 9:16 vertical output, 1080p and 4K upscaling, and significantly improves consistency in character identity, background, and objects. The feature is now available in the Gemini app, YouTube Shorts, Flow, Google Vids, and Vertex AI.
Zhipu, in collaboration with Huawei, has open-sourced the image generation model GLM-Image. The model completed its entire workflow, from data preprocessing to large-scale training, based on Ascend Atlas 800T A2 devices and the MindSpore AI framework.
Google Research published a paper in Science Advances detailing how NeuralGCM trains its AI modules by directly using NASA satellite precipitation observation data from 2001–2018.
Berkeley AI Research published a paper at NeurIPS 2025 proposing the Information-Driven Encoder Analysis Learning (IDEAL) framework.
ElevenLabs has released Scribe v2, targeting batch transcription, subtitling, and annotation scenarios. It improves stability and accuracy for long audio, supports context-aware Keyterm Prompting, automatic detection of 56 types of entities with precise timestamps, automatic language identification within a single file, intelligent speaker diarization, word-level timestamps, and non-speech event tagging. It also complies with multiple certifications including SOC 2. The product is now integrated into ElevenLabs Studio and API.
SGLang releases Model Gateway v0.3.1, with core upgrades including a Radix Tree / Cache-Aware routing engine: cache insertion throughput reaches 216.
The Tongyi Qianwen team has open-sourced the Qwen3-VL-Embedding and Qwen3-VL-Reranker model series, built upon Qwen3-VL, supporting mixed-modality inputs including text, images, screenshots, and videos.
The Baidu Baihe team completed deep adaptation and optimization of the inference engine for the DeepSeek R1 model on a Baidu Tianchi 32-card Kunlun Chip XPU super node, adopting a PD separation architecture with TP4+SP4 during the Prefill stage (extended to TP32+SP32 for long sequences).