DeepSeek V4.1 Flash in Token Factory: One Million Token Context

DeepSeek V4.1 Flash has been added to the catalog of Token Factory, Selectel's pay-as-you-go inference service. This October update of the open model from the Chinese laboratory DeepSeek features weights released on October 17 under the MIT license. The model operates via the same OpenAI-compatible API as the rest of the catalog, while all requests are processed within Russian data centers.

• The updated DeepSeek V4.1 Flash with open weights under the MIT license is now available in Selectel Token Factory.
• The 763B parameter MoE architecture activates about 16B parameters per token, delivering high speed and an efficient KV cache of 890 bytes per token in FP4.
• The model supports multimodality with a custom DeepSeek-ViT encoder and a one-million-token context window, outperforming predecessors in agentic tasks.

Architecture and Technical Features

DeepSeek is a Hangzhou-based laboratory whose open models have repeatedly disrupted API market pricing. While the base V4.1 Flash launched in September, the company refers to the 1017 version as a minor update. Architecturally, it is a Mixture of Experts (MoE) model with 763 billion total parameters, of which approximately 16 billion are active per generated token.

The primary engineering value lies in the highly efficient KV cache. In FP4 format, it occupies 890 bytes per token—four times less than V4 Flash. This exact approach enables effective memory scaling when handling a context window of one million tokens. Additionally, the model is multimodal, accepting images through DeepSeek's proprietary, scratch-trained vision encoder.

Performance and Use Cases

DeepSeek promises higher benchmark scores, faster task execution, and more concise responses. An 11-point performance boost on Terminal-Bench highlights significant improvements in terminal-based agent workflows. Furthermore, in automation and web development benchmarks, the lightweight Flash variant managed to outperform the heavier Opus model.

The new context window is 2.7 times larger than that of V4 Flash. For a typical agent query with 50,000 input tokens and 2,000 output tokens, costs drop significantly when leveraging the cache. While chats and simple classification tasks remain more cost-effective on standard V4 Flash, version 1017 is justified when maintaining entire code repositories or long documents in active memory.

Data Security and Integration Details

All requests remain securely inside the Russian infrastructure. According to the documentation, prompts and responses are neither stored nor used for model training; only technical consumption statistics are gathered. For projects dealing with personal data, this serves as a compelling argument for local inference over direct API usage.

Integration is seamless thanks to OpenAI API compatibility. Developers only need to update the base_url and model name in existing codebases, while DeepSeek recommends setting temperature and top_p parameters to 1.0. This update provides developers working on local infrastructure with extended context and vision capabilities without requiring architectural rewrites.