Shanghai AI Lab (Shanghai Artificial Intelligence Laboratory) has released the full version of Atria Dawn, an AI model designed to carry out complex work such as research and software development across multiple stages. The official API supports a context length of up to 512K tokens, double the 256K of the Preview version. This makes it easier to take on tasks that span multiple steps while referencing long documents and past work history. However, the full version's model weights have not yet been released, so its availability differs from the downloadable Preview version. Evaluating this update requires distinguishing between what can be used through the API and what can be run in your own environment.
512K-token API now available; full-version weights still pending
Atria Dawn is an agentic AI model built on GLM-5.2, a foundation model with 744 billion parameters. It uses a Mixture of Experts (MoE) architecture, which combines networks that each handle specialized processing, and aims to advance multi-step tasks using tools while keeping track of its surroundings and the progress of its work. It is positioned not merely as a model that answers questions, but as one that handles an entire workflow: analyzing a problem, designing a solution, implementing code, running experiments, and verifying results.
Pandaily's October 17 report noted that while the context length has expanded to 512K tokens, the model distribution table on GitHub lists 256K. The official API documentation states clearly that the available model ID "Atria-Dawn" has a 512K context length, and explains that this limit covers both input and generated output. In other words, not all 512K tokens can be used for reference material in the input. System instructions, conversation history, document excerpts, and results returned from tools all consume context, so room must be left for generating the response. While more history can be referenced during long-running research or code fixes, designing what to retain and what to organize or discard remains important.
Context length is also distinct from the number of tokens that can be generated in a single response. According to the official API documentation, the parameter that sets the output token limit accepts values from 1 to 65,536, but actual output length also depends on stop conditions and the remaining context capacity. The 512K-token support therefore does not mean a large volume of text can be generated at once; it should be understood mainly as an expansion of the amount of information that can be referenced.
Model weights are the set of numerical parameters determined through training, and they are required to run an AI model on your own servers or similar infrastructure. Based on the model distribution table in the official GitHub repository, the official API documentation, and the Preview version's model card, the availability as of October 17, 2026 is as follows.
| Item | Stated context length | Availability |
|---|---|---|
| Full-version API | 512K | Official documentation describes how to call it |
| Full version and full-version FP8 (distribution table) | 256K | "Coming soon" on both Hugging Face and ModelScope |
| Preview and Preview FP8 | 256K | Download links listed in the official repository |
The official API explicitly states 512K-token support, while the GitHub distribution table lists 256K, and no explanation is given for the discrepancy. It therefore cannot be assumed that 512K tokens will be available when the full-version weights, once released, are run in your own environment. The statement that the code and model weights will be released under the MIT license also does not mean the full-version files are already available for download.
Official benchmarks show gains on 17 of 20 items
According to the benchmark results published by Shanghai AI Lab, of the 20 items comparable with the Preview version, the full version of Atria Dawn improved on 17, stayed flat on 1, and declined on 2. Improvements are reported across a wide range of areas, including research, software development, tool use, and business support. However, performance did not rise on every item, which indicates that the update to the full version may have reduced some capabilities.
This tally is based on the "Evaluation Results" posted in the official GitHub repository as of October 17, 2026. It extracts the items for which figures are listed for both the full and Preview versions and compares the scores. Three items for which one of the results is not listed (DeepSWE, Harvey LAB, and OfficeQA Pro) are excluded from the tally. We also avoided treating unlisted results as zero and averaging scores that use different metrics. The main improved items, along with those that stayed flat or declined, are as follows.
| Benchmark | Preview | Full version | Score change |
|---|---|---|---|
| JobBench | 50.3 | 57.5 | +7.2 |
| WideSearch | 81.9 | 84.6 | +2.7 |
| PaperBench | 87.0 | 89.6 | +2.6 |
| SWE-bench Pro | 59.6 | 61.7 | +2.1 |
| Harbor-Index | 17.3 | 17.3 | 0 |
| BrowseComp | 92.5 | 91.3 | −1.2 |
| NL2Repo-Bench | 51.4 | 50.0 | −1.4 |
Note: The score change is the Preview value subtracted from the full-version value; it does not represent a rate of improvement.
The largest improvement is on JobBench, which evaluates tasks close to real-world work. The score rose 7.2 points, from 50.3 for the Preview version to 57.5 for the full version, and Pandaily also highlights it as a notable improvement. However, the same official comparison table lists GLM 5.3 at 58.2 and Claude Opus 5 at 68.0, both above Atria Dawn's 57.5. Improving on the previous version is clearly different from achieving the highest performance in a comparison that includes competing models.
On the other hand, BrowseComp fell 1.2 points, from 92.5 to 91.3, and NL2Repo-Bench fell 1.4 points, from 51.4 to 50.0. This shows that the update to the full version did not improve every capability uniformly. Nor can the fact that scores rose on 17 of 20 items be read to mean that performance will improve 85% of the time in real work. When deciding whether to adopt the model, it is more appropriate to pick the benchmarks closest to the work you want to delegate and focus on those results.
These figures are evaluation results published by the developer, not the results of independent verification by a third party. Whether conditions such as tool configuration, number of attempts, and compute usage are uniform across all compared models also cannot be confirmed from the published tables and explanations alone. The improvement on 17 items should therefore be understood strictly as a tally of score changes announced by the developer.
Putting the 512K context to use in real work
Shanghai AI Lab lists four main areas of use for Atria Dawn: Discovery, Creation, Delivery, and Cybersecurity. Discovery refers to collecting and organizing information and evidence and turning research questions into actionable plans, while Creation refers to developing software, machine learning systems, applications, and the like. Delivery covers work such as completing reports and presentations from documents, data, and design requirements, and Cybersecurity targets vulnerability verification, remediation, and re-verification in authorized environments. All of these are areas where work cannot be completed by answering a single question and requires carrying out multiple tasks in succession.
What these uses have in common is a workflow in which results obtained along the way are checked and the next step is decided accordingly. In software development, for example, it is not enough to write code; the results must be checked, any errors investigated and fixed, and the code tested again. In research, too, the outcome of each stage, from gathering information to examining hypotheses, running experiments, and analyzing results, must feed into the next decision. In its official GitHub repository, Shanghai AI Lab states as a development goal that such work should lead to "executable, verifiable, and reproducible outcomes." However, this describes the model's design philosophy and is not a guarantee that it can correctly complete every kind of task.
A large context of 512K tokens may be useful in such multi-step work. If more reference material, code, and past attempts can be held in a single context than before, it becomes easier to continue work while referring back to what happened along the way. But a longer context length is not the same as improved ability as an agent. Using the API alone does not automatically set up a code execution environment or integration with external tools; the actual tool execution is handled by the application or agent side, which needs a mechanism to return the results to the model. Only when the ability to hold a long history is combined with a mechanism for choosing appropriate tools and verifying execution results can the model be applied to complex work.
There are also constraints on input format. Atria Dawn accepts text only; it is not a multimodal model that can read images or PDF files directly and understand their contents. To analyze papers or technical reports, the necessary information must first be extracted as text and passed to the model. Even with the context length expanded to 512K, information contained in images and figures cannot be understood directly, so how documents are preprocessed will also affect the quality of the final output.
The Atria Dawn API supports three formats, Chat Completions, Messages, and Responses, and can be used from existing applications and AI agents. When considering adoption, it is important to measure the accuracy of responses and generated code, instruction-following, response time, and token usage using the actual work you intend to delegate. When using a long context in particular, it is necessary to confirm not just whether large amounts of information can be input, but whether the model can correctly reference the necessary information and deliver consistent results across the whole task.
Once the full-version weights are released, running the model in your own environment will become a concrete option alongside API access. However, that decision requires checking specifications such as the context length at distribution, the compute resources required, and inference speed. With this full version, the API context length has expanded to 512K tokens, and the developer's evaluations report improvements on many items. How far that performance carries over to real research and development work, however, remains to be verified. Along with the release of the full-version weights, accuracy and processing performance when handling long contexts will be key factors in judging Atria Dawn's practicality.