The Chinese model may be strong while the product around it feels unfinished

A Chinese model can be technically strong while the product used to reach it feels awkward, incomplete or incapable of ordinary work. That is not a contradiction. A foundation model is only one layer. The user encounters an application, file parser, tool system, storage service, regional API, account policy and export pipeline. Weakness in any one of those layers can make a capable model feel like an early chatbot.

Kimi, DeepSeek and Qwen illustrate three different product strategies. Kimi has invested heavily in an integrated consumer workspace for documents and spreadsheets. DeepSeek’s current public Files API is deliberately an image service for its Flash model rather than a general Office-document layer. Alibaba Cloud exposes powerful document services, but through specialised models, region-specific endpoints and capability matrices that builders must assemble correctly.[1][2][3][5]

Kimi is not missing the Office layer

Kimi Projects can carry files and instructions across related chats, reading project files when the model decides they are relevant rather than injecting every file into every turn. Its current documentation supports PDF, DOCX, XLSX, CSV, text, Markdown, code and image formats, with a 100 MB per-file ceiling and up to 50 files. Kimi Docs and Kimi Sheets add a higher-level workflow: edit an uploaded Word, PDF or spreadsheet and export a new Office-format artifact.[1][2]

That means a user who tried only a basic Kimi chat or an older model may have encountered a much narrower experience than the current specialised tools. The existence of those tools does not prove that every complex document will be preserved perfectly. It does show that “Kimi cannot work with files” is too broad a conclusion.

DeepSeek’s API exposes a narrower file primitive

DeepSeek’s current Files API stores reusable images for deepseek-flash. The supported formats are JPEG, PNG, GIF and WebP; a stored image can be as large as 64 MiB. The vision guide describes screenshots, charts and pictures. It does not document the Files API as a PDF, Word or spreadsheet parser. A developer sending a DOCX file to that endpoint is using the wrong interface rather than measuring the language model’s reasoning ability.[3][4]

A consumer app can still offer text extraction or document features through a separate preprocessing service. That distinction matters because the model may receive extracted text rather than the original file structure. Tables, tracked changes, formulas, page layout and embedded media can be lost before the model begins reasoning.

Qwen’s strength comes with product fragmentation

Alibaba Cloud’s Qwen-Long can reference up to 100 uploaded files and advertises a ten-million-token context, but the documented service is tied to the China Beijing region. It supports many common document formats and scanned PDFs. Its current guide also supports JSON-schema structured output. Builders should check function calling and search separately: a document-analysis endpoint, a schema-constrained response and an agent tool are different capabilities.[5][6]

Qwen-Doc-Turbo serves another job: document extraction and reconstruction across Word, spreadsheet, presentation and image inputs. Its limits and SDK path differ from Qwen-Long. This is capable infrastructure, but it creates a discovery problem. The user must know whether they need long-context question answering, document generation, multimodal extraction, translation or tool use before selecting the model and region.

The fair conclusion is not that Chinese AI is bad or that Western products are universally complete. Kimi now presents a substantial document workspace; DeepSeek’s API chooses a narrow image-file design; Qwen offers unusually large document services through a more fragmented cloud catalogue. The useful comparison begins with the work to be done and follows the entire path from file upload to editable output.

Report an error or suggest a correction