Frontier Models
DeepSeek V4.1 Flash Releases CED Architecture and Open Weights to Cut Frontier Costs
The new model activates far fewer parameters than its size suggests while adding native vision and a 1 million token context at sharply reduced API rates.
DeepSeek V4.1 Flash is a 552 billion parameter multimodal Mixture-of-Experts model that employs a Causal Encoder-Decoder architecture to activate only 8 billion parameters during prefill and 16 billion during decode while supporting up to 1 million token contexts and native vision.
The release replaces the prior V4 Pro flagship with a design that prioritizes efficiency without sacrificing results on agentic benchmarks.
DeepSeek trained the model from scratch on 45 trillion multimodal tokens and extended context length to the full 1 million tokens.
What background led to the V4.1 Flash release?
Earlier DeepSeek models required higher active parameter counts during both input and output stages.
The company identified memory and compute bottlenecks as barriers to wider deployment of large context models.
The new architecture family addresses those constraints while preserving multimodal capabilities.
How does the Causal Encoder-Decoder architecture operate?
The encoder processes input tokens with 8 billion active parameters while the decoder generates output with 16 billion active parameters.
This split allows the model to allocate compute asymmetrically across the inference pipeline.
The design also compresses the KV cache to approximately one quarter of previous high-bandwidth memory requirements and one eighth of previous solid-state drive requirements.
| Specification | Detail |
|---|---|
| Total Parameters | 552 billion |
| Active Parameters Prefill | 8 billion |
| Active Parameters Decode | 16 billion |
| Maximum Context Window | 1 million tokens |
| Maximum Output Length | 384 thousand tokens |
| License | MIT |
Which benchmark scores stand out for the model?
The model records strong results on reasoning and software engineering evaluations at the maximum reasoning effort setting.
It reaches 90.6 percent on Terminal-Bench 2.1 Pass@1 according to the same evaluation protocol.
What are the updated API pricing tiers?
Off-peak cache-miss input pricing stands at $0.15 per million tokens and output pricing at $0.60 per million tokens.
Peak rates are set at $0.30 per million input tokens and $1.20 per million output tokens.
The API endpoint uses the model name deepseek-flash with prior V4 Flash variants retired through compatibility routing.
What market and stakeholder implications follow from the release?
Developers can access competitive performance at a fraction of typical frontier model costs.
The MIT license and Hugging Face hosting enable fine-tuning and local inference without usage restrictions.
Enterprises gain a route to scale agentic applications where previous pricing limited experimentation.
Tests by multiple parties put V4.1-Flash ahead of V4-Pro on performance, cost, speed & total runtime. We're phasing out V4-Pro.DeepSeek
What timeline governs the rollout and future releases?
The weights appeared on Hugging Face on or around September 10 2026.
V4 Pro traffic begins routing to the new model at Flash rates on September 14 2026 at 04:00 UTC.
- September 10 2026: Weights released on Hugging Face under MIT license
- September 14 2026 at 04:00 UTC: V4-Pro requests route to deepseek-flash at Flash pricing
- Subsequent months: Planned launch of V4.1-Pro variant
- Ongoing: User feedback incorporated into further optimizations
The phased approach minimizes disruption for existing API users while expanding access to the new architecture.
Frequently asked
What active parameter counts apply during different inference stages?
The model activates 8 billion parameters for input prefill and 16 billion parameters for output decode under the Causal Encoder-Decoder design.
When will V4-Pro requests begin using the new model?
Routing of all deepseek-v4-pro requests to V4.1 Flash at the lower rates starts at 04:00 UTC on September 14 2026.
Where can developers obtain the open weights?
The weights are available on Hugging Face at the repository deepseek-ai/DeepSeek-V4.1-Flash under the MIT license.
Sources
- DeepSeek — The announcement covers the introduction of the model, its architecture, the routing of V4-Pro requests, and the decision to phase out the prior flagship.
- Hugging Face — The model card provides the 552B parameter count, CED architecture details, 1 million token context support, benchmark scores including 90.9 percent on GPQA Diamond, and MIT license information.
- DataCamp — The article lists the off-peak API pricing at $0.15 per million input tokens and $0.60 per million output tokens along with the corresponding peak rates.
- DeepSeek — 今天，我们正式发布 DeepSeek V4.1 Flash 模型... 552B 参数的 MoE 模型，采用了全新的 Causal-Encoder-Decoder 结构，输入激活只有 8B，输出激活 16B... deepseek-flash... V4.1 Flash 在性能、费用、速度、总用时等各项指标上已全面超越 V4 Pro