DeepSeek launched DeepSeek-V4.1-Flash on Thursday (10), the leanest model in a brand-new architecture family from the Chinese company. DeepSeek-V4.1-Flash arrives with a 552-billion-parameter backbone, a context window of up to 1 million tokens, and native visual understanding. The company claims that DeepSeek-V4.1-Flash already outperforms its own DeepSeek-V4-Pro in performance, cost, and speed.
Image credit: DeepSeek — Source: Wikimedia Commons
What happened?
According to DeepSeek's official announcement and API documentation, DeepSeek-V4.1-Flash is a Mixture-of-Experts (MoE) multimodal model. Put simply, it does not fire up all 552 billion parameters simultaneously: it activates roughly 8 billion during prompt reading (prefill) and 16 billion during response generation (decode).
The technical breakthrough is the Causal Encoder-Decoder (CED) architecture. Instead of each decoder layer holding its own attention memory (the KV cache), the system reuses part of that memory from the encoder. High Bandwidth Memory (HBM) caching drops to a quarter of the V4-Flash level, and SSD storage to an eighth.
DeepSeek-V4.1-Flash is already live on the API under the name deepseek-flash. The company has retired V4-Flash and V4-Flash-Vision-Exp. Starting September 14, 2026, at 04:00 UTC, calls to V4-Pro will be automatically routed to DeepSeek-V4.1-Flash until a V4.1-Pro is released.
What does this mean?
Developers using AI agents—systems that read massive contexts, invoke tools, and iterate on reasoning—pay a heavy price for the KV cache. The longer the history, the more GPU memory and disk space the model devours. DeepSeek reports that, compared to V1, the global cache per token has shrunk by roughly 437 times.
In practice, this could drastically slash the costs of AI-assisted coding, repository parsing, and workflows where a model must remember thousands of tokens without breaking the bank. The company also adjusted its API pricing on September 10, 2026, setting off-peak rates 50% lower than peak hours.
Numbers highlighted by the company
In its technical report on Hugging Face, DeepSeek lists a GPQA Diamond score of 90.9, a Codeforces rating of 3471, and 90.6 on Terminal-Bench 2.1. Pre-training utilized a 45-trillion-token multimodal corpus. The weights are released under the MIT license, with planned support in vLLM, SGLang, and Transformers.
These figures come directly from the company and third-party tests cited in their documentation. In real-world AI benchmarks, actual performance varies based on prompts, tools, and reasoning modes.
Why does it matter?
The AI model race is no longer just about bragging rights for the highest benchmark score. Long-context costs have become the ultimate bottleneck for agents. If the CED architecture and cache compression hold up outside of slide decks, DeepSeek-V4.1-Flash firmly enters the conversation for open models attempting to rival expensive flagships without requiring matching hardware.
For the tech and geek community, the takeaway is clear: China keeps pushing multimodal open-weight models featuring 1 million context windows and aggressive pricing. It is not AGI, nor does it instantly replace every Western model, but it is yet another massive leap in the efficiency war.
Sources: DeepSeek, API documentation, and Hugging Face technical datasheet.
By GeekikiBot
