DeepSeek has launched DeepSeek-V4.1-Flash, a new model with a 552-billion-parameter mixture-of-experts backbone, native vision, and a 1-million-token context window. This architecture is specifically designed to reduce the cost of repeatedly reading large contexts, making it highly efficient for tasks that involve extensive data reuse.
The model is priced to optimize cost-efficiency, especially during off-peak hours, where it costs $0.003 per million input tokens on a cache hit, $0.15 per million on a cache miss, and $0.60 per million output tokens. During peak hours, which are Monday through Friday from 01:00–04:00 UTC and 06:00–10:00 UTC, these rates double. This pricing structure is particularly beneficial for developers working on coding agents and other long-running workflows, where cache hits can significantly reduce costs.
DeepSeek’s peak hours are Monday through Friday from 01:00, 04:00 UTC and 06:00, 10:00 UTC. All other hours are considered off-peak, providing developers with a strategic window to optimize costs by scheduling workloads during these times.
A focus on cost-efficiency
The pricing structure is particularly relevant for developers working on coding agents and other long-running workflows. An agent repeatedly working against the same repository or conversation history may reread far more cached context than fresh context, making cache-hit pricing a critical factor in cost management. DeepSeek emphasizes that cache-hit charges can account for a significant portion of agent costs, highlighting the importance of optimizing for cache reuse.
A VentureBeat survey of 170 enterprises found that only 47% rigorously track AI compute cost and ROI, indicating a gap in cost management practices. Additionally, 12% have not addressed inference-memory limits, and 7% were unaware of the constraint, showing the need for better cost-tracking tools and awareness in the industry.
DeepSeek’s off-peak cached-input rate of $0.003 per million tokens is unusually low compared to competitors like OpenAI, which charges $0.40 for cached input on its GPT-5.6 Sol model. This makes DeepSeek-V4.1-Flash an attractive option for cost-sensitive applications, especially those with high cache-hit ratios.
Technical innovations
DeepSeek-V4.1-Flash employs a Causal Encoder-Decoder architecture, dividing its 40 Transformer layers into a 20-layer causal encoder and a 20-layer decoder. This design activates 8 billion parameters per token during input reading and 16 billion during output generation, optimizing efficiency without sacrificing performance.
Read Also: Databricks agents match top AI answer quality in half
The model incorporates Compressed Sparse Attention 2, hierarchical sparse indexing, and FP4 KV caching, reducing the global KV cache size to 890 bytes per token, approximately one-quarter of its predecessor. Persistent cache storage requirements are also reduced to roughly one-eighth, significantly improving memory efficiency.
Despite these optimizations, the model’s total footprint has increased from 284 billion to 552 billion parameters, making local deployment more resource-intensive. This trade-off between efficiency and size has sparked debate among developers, with some questioning whether the model remains “flash” in the traditional sense.
Benchmarks and performance
DeepSeek reports a 74.2 score on DeepSWE v1.1, slightly outperforming Claude Opus 5 and GPT-5.6 Sol. It also scores 88.1 on CyberGym and 54.8 on AutomationBench, demonstrating strong performance across various benchmarks.
Early third-party evidence suggests that V4.1-Flash can achieve 98% of GPT-6 Astra’s quality at 1.4% of its cost on everyday design requests, highlighting its cost-effectiveness. However, the model’s performance is workload-dependent, and developers must consider factors like cache-hit ratios and reasoning effort to maximize efficiency.
DeepSeek has retired V4-Flash and V4-Flash-Vision-Exp, routing their identifiers to V4.1-Flash. Starting September 14, 2026, calls to deepseek-v4-pro will also be routed to V4.1-Flash until V4.1-Pro is released. This transition has drawn criticism from developers concerned about regression testing, as changes to underlying models can impact existing workflows.
The launch coincides with DeepSeek’s reported preparations for a potential listing on Shanghai’s STAR Market, with a valuation of up to $75 billion. This strategic move shows the company’s growing commercial ambitions and the significance of V4.1-Flash in its product lineup.
