DeepSeek V4.1 Flash Arrives in Selectel's Token Factory Catalog

The main engineering feature is a frugal KV cache: in FP4 it takes 890 bytes per token, four times less than V4 Flash. This makes memory scale efficiently with a context of one million tokens, 2.7 times larger than V4 Flash. The model is multimodal, with images handled by DeepSeek-ViT, an encoder trained from scratch. DeepSeek promises higher benchmark results, faster task execution and shorter answers, which matters under per-token billing. The technical report shows an 11-point gain on Terminal-Bench, and in automation and web-development tests the lightweight Flash outperformed the heavier Opus.

Cost and data handling

For a typical agent request with 50,000 input and 2,000 output tokens, the cost is about 13 rubles without caching. If 40,000 input tokens hit the cache, such as a repeated system prompt and tool descriptions, it falls to about 5.8 rubles.

The practical rule: V4 Flash remains cheaper for chats, classification and bulk text processing, while version 1017 makes sense when a whole repository, a long conversation or scanned documents must stay in memory. All requests stay inside Russian infrastructure. According to the documentation, prompts and responses are not stored or used for model training; only technical consumption statistics are collected. For projects with personal data, that is likely the main argument for local inference over DeepSeek's direct API.

To connect, open Products → Token Factory in the control panel, create an API key and copy the model alias deepseek-v4.1-flash-1017 from the model catalog. Because the API is OpenAI-compatible, existing code only needs a new base_url and model name. DeepSeek recommends temperature and top_p of 1.0.