This observation covers technical changes disclosed and verified around 7:49 am Sydney time on 20 November to 7:49 am on 21 November 2026. Four items are included.
Tencent delivers the final Hy4 weights
Tencent Hy released the final Hy4, Hy4-FP8 and Hy4 base weights on 20 November, with download locations on Hugging Face, ModelScope and GitCode and deployment instructions for vLLM and SGLang. The Hy4 preview released on 28 August had already provided a 770-billion-parameter mixture-of-experts model with 49 billion active parameters per token and a context window of about one million tokens. The final model was released on 4 November; the new fact in this window is that the final instruction model, FP8 model and pretrained base model can all be downloaded and deployed independently. The current model card lists text, image and video understanding, with the same 770-billion-parameter language backbone and 49 billion active parameters, plus a 1.23-billion-parameter vision encoder and a 3.3-billion-parameter DFly2 draft model.
The weights use Tencent Hy License 1.0 rather than Apache 2.0. The licence permits modification and redistribution, but MaaS operators with more than US$50 million in worldwide revenue over any consecutive 12 months, or products with more than 100 million monthly active users, must obtain a separate written licence. Performance figures are mainly Tencent internal and developer evaluations, without a published complete independent replication. Tencent also lists remaining gaps against leading proprietary models in research-level mathematics, long-horizon coding, model self-improvement and advanced visual question answering.
Google opens Genie 4 through the Gemini API
Google first provided Genie 4 as a research preview to trusted testers on 27 October, opened Project Genie to eligible Google AI Ultra subscribers on 18 November, then made it available to developers on 19 November through the Gemini API's Interactions API and Google AI Studio. Developers can start from text, a reference image or both, then control a continuous interactive world with keyboard input and live text prompts. Google's specified output is 720p at up to 24 frames per second with synchronised native audio, and every output carries SynthID.
The model card explicitly describes the release as a preview. It says visual details and object permanence can drift even during short sessions, while complex physics, audio accuracy, legible text, spatial accuracy and long-term consistency remain limited. Google has disclosed only internal human and automated evaluations without complete scores that can be independently checked. As of 20 November, third-party reporting had not provided an independent benchmark or public replication.
OpenAI publishes the MedChemBench technical report
OpenAI formally published its MedChemBench introduction on 20 November together with a technical report dated 9 November. The retrospective multimodal benchmark contains 99 medicinal chemistry problems drawn from 117 papers and patents and includes 761 distinct molecular structures. It evaluates molecular reconstruction, scientific explanations, candidate ranking and selection, and numerical or categorical prediction. OpenAI reports that GPT-6.2 Astra earned 84.5 per cent of available credit at the xhigh reasoning setting shared across the evaluated models; the paper reports 83.3 per cent for GPT-6 Astra. These are developer evaluations and have not been independently replicated.
The report also limits the scope of the result. Tasks come from existing literature, so training-data exposure or memorisation cannot be excluded. Some open-ended design and synthesis tasks may penalise viable approaches that differ from reference answers. The single-turn, fixed-evidence protocol does not test tool use, additional experiments or iterative correction. The published material does not establish that the models can discover new drugs or improve real laboratory decisions.
OpenAI makes GPT-Live-1.1 available
OpenAI's API changelog records GPT-Live-1.1 as a feature released on 19 November, with a developer announcement on 20 November confirming API availability. It is a full-duplex multilingual voice model that can listen while speaking. Compared with GPT-Live-1, OpenAI reports better pronunciation, accent consistency and transcription, and the ability to delegate reasoning or tool-dependent work to a backend during a conversation. It is available through Live sessions; streaming and function calling are supported, while structured outputs and fine-tuning are not.
Voice sessions are billed by duration at US$0.05 per active minute. A WebRTC session can enter standby after 5 to 60 seconds of silence, keeping the connection open while inference pauses, at US$0.02 per minute. Backend model and tool use are billed separately. The improvement claims come from OpenAI, with no public third-party voice evaluation or independent replication available.
Sources
Tencent Hy, “Hy4”, version accessed 20 November 2026, https://github.com/Tencent-Hunyuan/Hy4
Tencent, “Tencent Releases and Open-Sources Tencent Hy4 preview”, 28 August 2026, https://www.tencent.com/tencent-releases-and-open-sources-tencent-hy4-preview/
ITHome, “Tencent Hy4 officially open-sourced, completing the Hunyuan LLM family release”, 20 November 2026, https://www.ithome.com/1/029/841.htm
Google, “Genie 4 is now available on the Gemini API”, 19 November 2026, https://blog.google/innovation-and-ai/technology/developers-tools/genie-4-in-gemini-api/
Google DeepMind, “Genie 4 – Model Card”, 18 November 2026, https://deepmind.google/models/model-cards/genie-4/
Dirty Static, “Genie 4 API: 720p Worlds At 24 Frames”, 20 November 2026, https://dirtystatic.com/articles/genie-4-api-720p-worlds/
OpenAI, “Can AI help scientists identify promising drug candidates?”, 20 November 2026, https://openai.com/index/introducing-medchembench/
OpenAI, “MedChemBench: Evaluating complex multimodal reasoning in medicinal chemistry”, 9 November 2026, https://cdn.openai.com/pdf/09c4ec60-a32e-4652-837b-1811368a3bb0/medchembench-paper.pdf
OpenAI, “Changelog”, 19 November 2026, https://developers.openai.com/api/docs/changelog
OpenAI, “GPT-Live-1.1 Model”, version accessed 20 November 2026, https://developers.openai.com/api/docs/models/gpt-live-1.1
OpenAI Developer Community, “GPT-Live-1.1 has arrived in the API”, 20 November 2026, https://community.openai.com/t/gpt-live-1-1-has-arrived-in-the-api/1418537