AI Technology Observations | 3 November 2026, 07:00

This observation covers the approximately 24-hour period from 7:47 am on 2 November to 7:47 am on 3 November 2026 in Australia/Sydney. Two formally released and verifiable technical changes are included.

LongCat-2.5 advances from preview to formal API service

LongCat confirmed in its platform change log that LongCat-2.5 became available on 2 November. Relative to the LongCat-2.5-Preview service opened on 25 September, the new state is the formal model entering the online chat and API platforms. The model accepts text and images, represents visual signals as multi-level discrete tokens and processes them with text in a unified format. It targets visual coding, creative design, 3D tasks, complex codebases and complete project execution. Official material gives a scale of 1.6 trillion total parameters with about 48 billion activated per token. The API documentation lists a 1,048,576-token context limit, with separate limits for output and thinking budgets.

The developer says training incorporates more long-horizon trajectories, large-scale reinforcement learning and reproducible tasks derived from real agent failures. Its W4A4 inference system quantises both expert weights and activations for domestic AI accelerators. Throughput, latency and benchmark figures in the technical post are developer tests; some comparison results come from public reports or official leaderboards, and no public independent reproduction is available. The confirmed delivery is the hosted service. At the observation cut-off, LongCat's GitHub and Hugging Face organisations did not provide LongCat-2.5 weights, and reporting described a weight release as a subsequent plan.

Hydrogen-1.0 releases scientific-model weights in three deployment versions

Deep Principle formally released Hydrogen-1.0 on 2 November. The model applies scientific post-training to GLM-5.3 for chemistry, materials science and life sciences. Its training description covers supervised fine-tuning, reinforcement learning, domain knowledge, laboratory records and discovery trajectories that include failed attempts. The public artefacts include unquantised BF16 weights, an NVFP4 version for NVIDIA Blackwell and an MXFP4 version for AMD Instinct. The model weights use the MIT licence, while supporting code and documentation written by the team use Apache-2.0.

The public model card lists about 753 billion parameters, a one-million-token context window and a maximum output of 128,000 tokens. The team reports testing the quantised versions on eight B300 GPUs and on AMD MI350X hardware. Complete training datasets and training recipes are not released with the weights. Results on FrontierScience-Research, GPQA Diamond and internal scientific tasks are developer evaluations without third-party reproduction. The official material also records Hydrogen-1.0 trailing GPT-6 Astra at its highest thinking setting on NMR structure elucidation. The model card requires scientific claims to be checked by specialists, primary literature, independent calculation or experiment, and warns that long-context capacity does not guarantee reliable use of every detail.

Sources

LongCat, Introducing LongCat-2.5, 1 November 2026, https://longcat.ai/blog/longcat-2.5

LongCat API Docs, LongCat API Platform Change Log, 2 November 2026, https://longcat.chat/platform/docs/ChangeLog.html

TechNode, Meituan launches LongCat-2.5 multimodal AI model for extended coding tasks, 2 November 2026, https://technode.com/2026/11/02/meituan-launches-longcat-2-5-multimodal-ai-model-for-extended-coding-tasks/

Deep Principle, Introducing Hydrogen-1.0, 2 November 2026, https://blog.deepprinciple.com/introducing-hydrogen-1-0-zh/

Hugging Face, deep-principle/Hydrogen-1.0-MXFP4, 2 November 2026, https://huggingface.co/deep-principle/Hydrogen-1.0-MXFP4

MarkTechPost, Deep Principle Releases Hydrogen-1.0: A 753B Open-Weight Model for Scientific Discovery, 2 November 2026, https://www.marktechpost.com/2026/11/02/deep-principle-hydrogen-1-0-open-weight-scientific-model/