# DeepSeek V4.1-Flash MoE Model Undercuts V4-Pro on Benchmarks and Cost

> The 552B-parameter release ships with native vision, MIT weights, 1M context and Flash-tier pricing while routing legacy Pro traffic to the new endpoint from mid-September.

*Published 2026-09-12 · By Marcus Vance*

DeepSeek V4.1-Flash is a 552B-parameter multimodal MoE model with native image understanding, a 1M-token context window, and MIT-licensed weights.

The release of DeepSeek V4.1-Flash marks a deliberate pivot toward asymmetric efficiency in frontier model design. The new model activates far fewer parameters than its total count would suggest, yet delivers stronger results on agent benchmarks than the company's own V4-Pro. This combination of performance and reduced resource use arrives with open weights and competitive API rates.

DeepSeek positions the model as the smallest member of a new architecture family. Legacy V4-Flash endpoints now point to V4.1-Flash, and V4-Pro traffic will begin routing to the Flash variant on September 14, 2026, at Flash pricing until the updated Pro model arrives. The strategy prioritizes immediate availability over waiting for the higher-tier release.

## What background prompted the shift to an efficient MoE variant?

Earlier DeepSeek releases emphasized scale through larger active parameter counts during both prefill and decode stages. Internal evaluations showed that selective activation could preserve or improve accuracy on software engineering and terminal tasks while lowering inference costs. The company therefore introduced the causal encoder-decoder MoE design to balance capability with operational economics.

The move also aligns with broader industry interest in models that support long contexts and multimodal inputs without proportional increases in memory or compute. By releasing under an MIT license, DeepSeek enables wider experimentation and fine-tuning by researchers and developers who previously faced restrictive terms.

## What technical specifications define the V4.1-Flash architecture?

The model employs a causal encoder-decoder structure that processes input with 8B active parameters and output with 16B active parameters. This split contributes directly to the reported KV cache reductions, requiring only one-quarter the high-bandwidth memory and one-eighth the solid-state drive storage of previous generations. The backbone contains 552 billion parameters in total.

Native image understanding is built into the architecture rather than added through adapters. The context window reaches 1 million tokens, supporting extended reasoning chains and large multimodal documents in a single pass. These choices reflect an emphasis on practical deployment constraints alongside benchmark performance.

Core specifications of DeepSeek V4.1-Flash drawn from official model documentation.SpecificationValueTotal Parameters552BActive Parameters (Input/Prefill)8BActive Parameters (Output/Decode)16BContext Window1M tokensArchitectureCausal Encoder-Decoder MoEMultimodal CapabilityNative image understandingLicenseMITKV Cache HBM Requirement1/4 of prior generationKV Cache SSD Requirement1/8 of prior generation

## How does V4.1-Flash perform on established agent benchmarks?

Evaluation results place the model at 90.6 on Terminal-Bench 2.1 under pass-at-one with maximum reasoning effort. It also reaches 74.2 on DeepSWE v1.1 for resolved tasks under identical evaluation conditions. These figures come from standardized testing protocols documented alongside the model weights.

The scores reflect performance in environments that require both code generation and execution feedback loops. Multiple independent parties reportedly confirmed advantages in speed and total runtime when compared with the V4-Pro predecessor on the same tasks.

## What market and stakeholder implications follow from the release?

API access carries peak pricing of $0.30 per million input tokens and $1.20 per million output tokens, with off-peak rates halved. The structure lowers barriers for high-volume agent workloads that previously required more expensive endpoints. Off-peak incentives further encourage flexible scheduling of non-urgent inference jobs.

Enterprise users gain an open-weight option that supports on-premises or private-cloud deployments without additional licensing fees. The reduced KV cache footprint translates into lower infrastructure spend for teams managing long-context or multimodal sessions at scale.

- Audit current V4-Pro traffic volumes and estimate savings under Flash routing.
- Validate benchmark parity on internal agent suites before the September 14, 2026 migration.
- Assess KV cache memory reductions in staging environments with 1M-token contexts.
- Plan fine-tuning experiments using the MIT-licensed weights hosted on Hugging Face.
- Schedule off-peak batch jobs to capture the 50 percent rate discount.

## What reactions have surfaced from the company and observers?

> 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

The statement signals a clear preference for the new architecture in production routing. Observers interpret the move as evidence that selective parameter activation can outperform denser configurations on targeted workloads without sacrificing multimodal capabilities.

## What developments are expected next from DeepSeek?

The company has indicated that V4.1-Pro remains on the roadmap, though no launch date has been confirmed. In the interim, all new traffic and legacy Pro requests will operate through the Flash endpoint at the lower rate structure. Continued updates to the technical report on Hugging Face are anticipated as additional evaluations complete.

Community adoption will likely center on the open weights for domain-specific adaptation. The combination of long context, native vision, and reduced inference cost positions the model for expanded use in research and commercial agent systems over the coming months.

## Sources

1. [The model scores 90.6 on Terminal-Bench 2.1 and 74.2 on DeepSWE v1.1 with 552B backbone parameters, causal encoder-decoder architecture, 8B input and 16B output activation, 1M context, and MIT license.](https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash)
2. [The official announcement states that tests show V4.1-Flash ahead of V4-Pro on performance, cost, speed and total runtime, with plans to phase out V4-Pro and details on pricing and routing changes.](https://api-docs.deepseek.com/news/news260910)
3. [The model introduces native visual understanding, 552B-parameter MoE design with 8B active for input and 16B for output, KV cache reductions to 1/4 HBM and 1/8 SSD, and live API access as deepseek-flash.](https://www.deepseek.com/en/news/deepseek-v4-1-flash/)
4. [DeepSeek launched V4.1-Flash, a 552B-parameter multimodal MoE with native image understanding, 1M-token context window, 8B/16B active parameters (causal encoder-decoder architecture), MIT license, and pricing of $0.30…](https://hasbeenreleased.com/models)

---
Source: https://aiintelreport.com/frontier-models/deepseek-v4-1-flash-moe-release
Index: https://aiintelreport.com/llms.txt · Full text: https://aiintelreport.com/llms-full.txt