Kimi scores 96.2% on a 1C database analytics test and emerges as a fallback if Claude access is blocked

1C developer Alexander Novikov lost access to Claude: his account was blocked in the middle of a paid period, and the remaining balance was not refunded. Worried the same could happen with GPT, he decided to line up a replacement among Chinese models in advance, and Kimi was the first he tried. The reason is simple: the subscription can be paid in rubles with a Mir card, and no VPN is needed. He ran the test on the same demo database and under the same rules he had used earlier for ChatGPT, Claude and Alice AI for Business. Kimi's result: a 96.2% match with the reference.

How the test works and what payment looks like

The Kimi Pro subscription cost 3300 rubles. The author separately warns that the Russian internet is full of links to resellers, so it is worth checking that you are paying on the official site. The payment went through normally.

The setup is simple: the Kimi desktop app for Windows with an English interface. Authorization goes through the website, after which the app links to the account. A COM connection did not work because the database was on a different computer, so Novikov used the 1C gateway Yellow Bridge.

The database is the Trade Management 11 demo, edition 11.5.25.112, with a pre-prepared reference. The ten questions cover reference lists, documents, stock balances, mutual settlements and tasks with multiple conditions.

Results by question

Kimi fully solved nine of the ten questions. It correctly counted warehouses and items without groups, listed all five organizations, and returned the five most recent posted sales documents with matching numbers, dates and amounts. It also correctly found customer orders for August 2022, the LED Camelion 1,5 stock balance at the right warehouse, and the debts of Assol and Altair. The model split the January and February 2023 totals by RUB and USD. On the task about stock of 1000 or more with no sales in 2023–2024, it found all six items with no extras.

The weak spot was the question about the top 3 products by sales in February 2023. The products were named correctly, but two of the amounts were wrong, hence 50%. The author links the error to the currency split. After a refined prompt, the figures matched the reference. The takeaway: for complex conditions, it is better to give a detailed query from the start.

On speed, the model made 24 calls to the database, and the queries took about 45 seconds. That is not the total dialogue time: query preparation is not included.

Writing to 1C and development

Kimi supports remote MCP servers over HTTP. If a write tool is connected and allowed, the model can not only read but also change data. According to the author, this is an advantage over Alice AI for Business, which for now can only read. For production databases it cuts both ways: write access should be enabled deliberately and only in a test environment.

For the development test, Novikov asked the model to take an extension from 1C, modify it and save it. The change was of medium complexity, and Kimi handled it. But the author promises a verdict on whether the model works as an everyday development tool later: the test focused mainly on analytics.

Limitations and practical takeaway

The test is narrow: one demo database, ten questions and one author. It is not a benchmark and not a substitute for independent evaluation, but a practical slice of a 1C analyst's tasks. For development there is not enough data yet, and the result depends on prompt quality.

The practical takeaway: Kimi looks like a workable safety net in case access to Western models is cut off. For analytics on a standard 1C configuration it is already usable, and the author considers it quite capable of writing code. For Russian teams, the result also matters for infrastructure reasons: payment in rubles, connection through a gateway and MCP support make it possible to build the model into a workflow without workarounds.