DeepSeek Launches V4.1 Flash: 552B Parameters, Open Source, and $0.15 per Million Tokens

Chinese company DeepSeek has introduced V4.1 Flash, a new multimodal language model built on an MoE architecture featuring a massive total capacity of 552 billion parameters and a context window of up to one million tokens. The model is optimized for software development, autonomous agent systems, and coding, with its open weights published on Hugging Face under the permissive MIT license.

Cost-Efficient Causal Encoder-Decoder Architecture

Despite its massive total size, the model relies on a Mixture of Experts (MoE) architecture with a Causal Encoder-Decoder mechanism, drastically reducing computational requirements. According to the model page on Hugging Face, only specific blocks of the neural network are active at any given time:

• When reading long documents, code, or conversation history, only about 8 billion parameters are engaged;
• When generating responses and writing code, a 16-billion parameter block is activated.

This setup is ideal for agentic tasks where the algorithm needs to continuously analyze huge amounts of input data before outputting a short command or code snippet. V4.1 Flash also natively processes images, allowing it to be used in multimodal pipelines.

Benchmark Results and Record-Low API Pricing

In DeepSeek’s internal tests, the new model performed strongly across specialized coding and agentic benchmarks, scoring 74.2 on DeepSWE v1.1, 88.1 on CyberGym, and 54.8 on Automation-Bench. On developer charts, V4.1 Flash outperforms GPT-5.6 Sol and Claude Opus 5 in several test cases. However, independent benchmark results on general suites have not yet been published, making head-to-head comparisons premature.

A key advantage of the new model is its combination of open weights and unprecedentedly low cloud access costs. According to the official pricing list on TokenCost, off-peak API usage costs just $0.15 per million input tokens and $0.60 per million output tokens.

As Reuters notes, this presents a compelling offer for developers of autonomous AI systems: an open-source model featuring a massive context window and low token costs is ideally suited for running parallel agents at scale for code refactoring, web automation, and scientific computing.