DeepSeek has launched its V4.1-Flash AI model, introducing a 552-billion-parameter Mixture-of-Experts architecture that dramatically reduces key-value cache memory requirements to one-quarter of its predecessor's footprint. By activating only 8 billion parameters during prefill and 16 billion during generation, the model achieves high-speed inference at a fraction of the cost of Western rivals. The release has triggered a mandatory API transition routing legacy V4-Pro traffic to V4.1-Flash by September 14, 2026, while sparking investor concerns over near-term semiconductor memory demand.
V4.1-Flash architecture improvements
- ▪DeepSeek-V4.1-Flash was trained from scratch on a multimodal corpus containing 45 trillion tokens, with native image and text processing capabilities.
- ▪DeepSeek-V4.1-Flash features a Causal Encoder-Decoder architecture that activates 8 billion parameters during input prefill and 16 billion parameters during text generation.
- ▪DeepSeek launched DeepSeek-V4.1-Flash on September 10, 2026, as the smallest model in its new architecture family.
- ▪DeepSeek-V4.1-Flash incorporates 196 billion N-gram parameters in a conditional memory module called Engram to store implicit knowledge and reduce compute requirements.
- ▪DeepSeek-V4.1-Flash is built on a 552-billion-parameter Mixture-of-Experts framework, which is close to double the 284 billion parameters of the previous V4-Flash.
KV cache memory reduction
- ▪DeepSeek-V4.1-Flash introduces Compressed Sparse Attention 2, which assigns Transformer layers to Full, Reindex, or Reuse modes to share cache and avoid redundant storage.
- ▪DeepSeek-V4.1-Flash eliminates persistent sliding-window attention storage on SSDs, reducing persistent key-value cache storage requirements to roughly one-eighth of the previous generation.
- ▪DeepSeek-V4.1-Flash reduces the global key-value cache footprint to 890 bytes per token, which is approximately one-quarter of the memory required by the previous V4-Flash.
- ▪DeepSeek-V4.1-Flash stores its main global key-value cache in 4-bit floating-point format, which is a reduction from the 8-bit format used in the previous generation.
Agent workload cost economics
- ▪DeepSeek-V4.1-Flash allows users to adjust its reasoning effort using an integer setting from 1 to 100, which trades off compute costs against accuracy.
- ▪A VentureBeat Pulse Research survey from July 2026 found that only 47% of 170 surveyed enterprises rigorously track AI compute cost and ROI.
- ▪DeepSeek-V4.1-Flash's architectural efficiency reduces the need for high-bandwidth memory, which contributed to a 3% drop in Samsung Electronics and SK Hynix shares on September 11, 2026.
API pricing structure
- ▪DeepSeek-V4.1-Flash peak API rates are double its off-peak rates, rising to $0.006 per million cached input tokens, $0.15 per million uncached input tokens, and $0.60 per million output tokens.
- ▪DeepSeek-V4.1-Flash off-peak API rates are priced at $0.003 per million tokens for cache hits, $0.075 per million tokens for cache misses, and $0.30 per million tokens for outputs.
- ▪DeepSeek's peak API pricing windows run Monday through Friday from 01:00 to 04:00 UTC and from 06:00 to 10:00 UTC.
Mandatory endpoint transition timeline
- ▪Starting September 14, 2026, at 04:00 UTC, DeepSeek will automatically reroute all API traffic directed to the deepseek-v4-pro endpoint to DeepSeek-V4.1-Flash.
- ▪DeepSeek has retired the V4-Flash and V4-Flash-Vision-Exp models, temporarily routing their legacy API identifiers to DeepSeek-V4.1-Flash.
Benchmark performance comparisons
- ▪DeepSeek-V4.1-Flash scored 90.6% on Terminal-Bench 2.1, outperforming OpenAI's GPT-5.6 Sol at 88.8% and Moonshot AI's Kimi K3 at 88.3%.
- ▪DeepSeek's technical report notes that DeepSeek-V4.1-Flash's DeepSWE v1.1 score varied from 65.5% to 74.2% depending solely on the agent evaluation harness used.
- ▪On the OpenDesign Arena benchmark, DeepSeek-V4.1-Flash scored 81.2 out of 100 on design tasks at a cost of $0.023 per design, compared to OpenAI's GPT-6 Astra which scored 82.7 at a cost of $1.61.
- ▪On the DeepSWE v1.1 software engineering benchmark, DeepSeek-V4.1-Flash scored 74.2%, slightly ahead of Claude Opus 5 at 74.0% and GPT-5.6 Sol at 73.0%.
Debatable claims
- ▪Using model distillation to train competitive AI models is ethically unacceptable
- ▪Western enterprises should avoid integrating AI models from Chinese companies
Story comments
Loading comments…