DeepSeek launches V4.1-Flash with asymmetric prefill, compressed KV cache, and open weights (2026)

DeepSeek released DeepSeek-V4.1-Flash on September 10, 2026, introducing an asymmetric 552-billion-parameter Mixture-of-Experts architecture with native visual understanding and a one-million-token context window. The model is live immediately on the DeepSeek API under the model identifier deepseek-flash, while full model weights are available on Hugging Face under an open-source MIT license. 1 2 3

What launched

Signal	Confirmed detail	Action window
Asymmetric Causal Encoder-Decoder	The 40-layer backbone uses a 20-layer causal encoder and a 20-layer decoder, activating only 8 billion parameters per token during prefill and 16 billion during decode across 552 billion total parameters. 2 4	Profile latency and memory usage on prompt-heavy agentic workflows where prefill costs dominate total execution time.
Quadrupled KV cache compression	Compressed Sparse Attention 2 and SWA Bounded Replay reduce the global key-value cache footprint to 890 bytes per token, requiring roughly one-quarter of the GPU memory and one-eighth of the SSD storage needed by DeepSeek-V4-Flash. 1 2	Increase batch sizes or concurrency limits on existing serving infrastructure without adding accelerator nodes.
Native multimodal pre-training	Visual understanding is built directly into the base language model via DeepSeek-ViT and 2D-RoPE across 45 trillion tokens, replacing the prior experimental vision wrapper. 2	Consolidate visual and text tool-calling agents into a single pipeline without separate vision endpoint calls.
API unification and deprecation roadmap	The API endpoint `deepseek-flash` is generally available. DeepSeek has retired `deepseek-v4-flash` and `deepseek-v4-flash-vision-exp`, while all requests to `deepseek-v4-pro` will automatically route to V4.1-Flash at Flash pricing starting September 14, 2026. 3 5	Update production model strings to `deepseek-flash` before September 14 to avoid unexpected routing behavior.
Open-source weights and tooling	DeepSeek released full BF16 and FP8 weights on Hugging Face under the MIT license, alongside prompt-encoding libraries and minimal inference code. 2	Verify tokenizer configs and use the official `deepseek-recipe` toolkit for local inference setup.

Efficiency gains and agentic benchmarks

DeepSeek adjusted API pricing to reflect the reduced serving footprint. At peak hours, input tokens cost $0.006 per million on cache hits and $0.30 per million on cache misses, with output tokens billed at $1.20 per million; off-peak usage carries an automatic 50% discount. 3

DeepSeek-V4.1-Flash agentic benchmark performance against frontier models. — *Official evaluation results on Terminal-Bench 3.0, DeepSWE v1.1, CyberGym, and Automation-Bench at maximum reasoning effort. 2*

At maximum reasoning effort, DeepSeek reports competitive marks against larger frontier models. The model scores 74.2% on DeepSWE v1.1, 90.6% on Terminal Bench 2.1, 88.1% on CyberGym, and 31.8% on Agent's Last Exam, while achieving a 3471 Codeforces rating. 2

Why it matters

DeepSeek-V4.1-Flash signals a deliberate shift from raw parameter scaling toward architectural memory efficiency. By cutting the key-value cache footprint to 890 bytes per token and activating only 8 billion parameters during prefill, DeepSeek makes long-horizon agent loops substantially cheaper to serve at scale. DeepSeek's planned deprecation of its own flagship V4-Pro endpoint underscores that efficiency advantage: the lab is replacing its higher-priced model with a faster, cheaper alternative. Development teams can migrate to deepseek-flash immediately for lower token costs, while self-hosting teams should review cluster requirements before staging multi-node deployments.