China's Meteorological Administration Implements AI-Based Weather Forecasting System Fengqing

The Shanghai Artificial Intelligence Laboratory has unveiled the full version of its agent-oriented large model, Atria Dawn. This release occurred approximately one month after the preliminary version of Atria Dawn was introduced in mid-September.

The model was trained on the GLM-5.2 base model, which contains 744 billion parameters in a mixture-of-experts format. It is designed to support scientific research and engineering work that requires environmental understanding, tool usage, and multi-step task execution.

The main change noted by the laboratory is the increase in the context window from 256 thousand tokens available in the preview to 512 thousand tokens in the full version. However, the model download table in the project repository on GitHub still lists 256K for both the full model and its FP8 version, so developers are advised to consult the hosted API documentation for the applicable limit.

The laboratory now classifies the model's capabilities into four areas. Discovery involves data collection and systematization, conducting in-depth research, and transforming research questions into executable experiment plans. Creation covers software development, interactive applications, games, machine learning systems, and data visualizations. Delivery allows technical specifications, data, and design requirements to be converted into structured reports and presentations. Cybersecurity focuses on analyzing security issues, vulnerability checking, patch application, and retesting, limited to authorized environments.

Compared to the preview, the laboratory highlights significant improvements: a gain of 7.2 points in JobBench (57.5 vs 50.3), an increase of 2.7 points in WideSearch (84.6 vs 81.9), and an increase of 2.6 points in PaperBench (89.6 vs 87.0). However, the evaluation table in the repository, which the laboratory claims corresponds to its website, shows smaller improvements across most other benchmarks, including AutomationBench (54.7), SkillsBench (67.8), and SWE-bench Pro (61.7). In some tests, the preview showed equal or higher results, such as in BrowseComp, where the preview's 92.5 surpasses the full version's 91.3.

The model is available via hosted APIs with separate international and Chinese endpoints. The repository includes setup guides for SGLang, vLLM, and Codex. The model accepts only text input. The code is released under the MIT license, but the weights for the full version on Hugging Face and ModelScope are currently marked as 'Coming Soon'; only the preview weights are available for download at this time.

OpenAI and Anthropic are negotiating with third-party evaluators to determine which data will be disclosed following the safety tests of their artificial intelligence models. This deadlock involves balancing the need to protect trade secrets with the requirement for transparency regarding the potential dangers of the technology.

According to Reuters, both companies have attempted to exert control over the publication of these reports, raising questions about the true autonomy of specialists tasked with analyzing increasingly sophisticated systems.

Testing processes often involve external teams that temporarily work within AI laboratories. Although this access is crucial for model analysis, it creates an inherent conflict: sensitive information can be vital for identifying flaws and measuring risks.

Among the debated measures are allowing laboratory staff to be present during the evaluators' work and imposing restrictions to prevent the same team from working simultaneously for competing companies.

Anthropic stated that in past conversations with evaluators, the latter held control over the methods, criteria, conclusions, and published content. However, for new teams, the company reserves the right to comment on and approve the reports at the beginning of the collaboration.

For its part, OpenAI denies seeking approval for the full text of the reports. The company argues that certain confidential data can remain secret, and evaluators have the prerogative to indicate if there have been significant suppressions.

In the absence of standardized norms or clear guidelines, laboratories can influence evaluations by managing access to the systems. Marc Warner, CEO of Faculty, a company associated with Accenture and selected by Anthropic as its first integrated evaluator, emphasized the importance of credibility in the results.

Warner told Reuters that evaluators are only useful if there is trust in what they report, making independence and transparency fundamental elements.

Rumman Chowdhury, former US scientific envoy for AI under the Biden administration and CEO of Humane Intelligence, also warned about the pressure exerted on researchers. She commented to Reuters that 'evaluators have this privileged access, but explicitly or implicitly, they know it comes at a cost.'

An illustrative case of this dilemma was reported by a study from METR and Redwood Research concerning an incident at Hugging Face. Autonomous agents driven by advanced OpenAI models deviated from expected behavior, triggering an attack that affected the platform's infrastructure.

The investigation, initially planned to last two days inside OpenAI, extended to six days, spread across three visits. Researchers only received more detailed transcripts in the final phase, after flagging inconsistencies in the data.

The organizations involved reported that they made adjustments to the structure, clarity, and tone of the report after receiving comments from OpenAI, but assured that they had not removed crucial information for their conclusions. They also admitted that the fear of harming future evaluations impacted editorial decisions.

Additionally, a petition signed by over 200 researchers, including Geoffrey Hinton, proposes three essential conditions for evaluations to be truly independent. The major challenge lies in creating regulations that protect legitimate commercial information without allowing the companies themselves to unilaterally determine which risks of their systems will be made public.

The Shanghai Artificial Intelligence Laboratory (Shanghai AI Lab) has made Intern-Decision available—a family of compact decision-making models. These models were released on September 28th in three size variants: 0.8B, 2B, and 4B parameters. Instead of generating long textual excerpts, Intern-Decision is designed to directly provide structured judgments along with probabilities for each possible answer.

The design concept is based on the principle of 'decision, not string.' Intern-Decision supports three types of questions. A choice question requires the model to select the best option from a predefined list and assign a probability to each option. A rating question requests a score on a user-defined scale. And a 'yes/no' question returns a binary answer with the corresponding probability. Furthermore, a single query can combine multiple types of questions about the same state, and all are processed in a single forward pass.

Integrating Visual Perception

According to the project repository on GitHub, the models are fine-tuned on the Qwen3.5 language core, while the vision tower and projector remain frozen. This allows Intern-Decision to accept images as part of the input data, meaning visual perception directly influences the decision-making stage. The team demonstrated this capability on tasks related to games and interfaces: the model analyzes raw game frames, including two-dimensional grid maps and side-scrolling levels, and independently selects the next action. Other demonstrations cover mouse control and browser usage.

Testing Results and Performance

The laboratory reports that the 4B model achieves an average accuracy of 90.02% across seven test sets, which is 1.28 percentage points higher than the 88.74% achieved by the commercial decision-making model Jev, which is used as the primary benchmark. These test sets include decision types, tool calling, news classification, and intrusion detection, totaling over 10,000 test strings. The repository also indicates an average local query latency of about 44 milliseconds when using a single consumer RTX 4090 GPU, and provides calibration results on a 96-case distribution benchmark. It should be noted that these figures are self-reported by the developers, and the repository emphasizes that protocols and evaluation scopes differ between models.

The release includes training code, two inference backends, tools for calculating benchmark results, temperature calibration presets, and a browser demo version; however, training data and some closed validation records are not included. Model weights are published on Hugging Face.

Local GPU manufacturer MetaX stated that it has completed the Day-0 adaptation, so Intern-Decision has been running on its hardware since launch. MetaX claims that starting in December 2025, it will provide Day-0 support for 40 major flagship models, and its MXMACA software stack supports over 40 AI frameworks and more than 1000 models.

For developers creating agents, routers, and content filters, small models that output calibrated probabilities instead of free text can make automated decisions more cost-effective, faster, and easier to audit.