DeepSeek V4.1 Flash: What Changed From V4 Flash and V4 Pro

DeepSeek-V4.1-Flash is a new open-weight, multimodal model released on September 10, 2026. Its headline is not simply model size: DeepSeek is pairing a 552B-parameter mixture-of-experts system with sparse activation and a dramatically smaller KV cache to make very long-context inference more practical. 1735

The practical model-lineup change

For developers using DeepSeek’s API, V4.1 Flash is the current Flash offering, called with the deepseek-flash model name. It adds native visual understanding alongside text support. 1517

The older V4 Flash and experimental V4 Flash Vision Exp models have been retired. DeepSeek says the legacy identifiers deepseek-v4-flash and deepseek-v4-flash-vision-exp are still accepted temporarily, but requests are served by V4.1 Flash and billed at Flash rates. 1523

V4 Pro is different. Early transition reporting said its traffic would be redirected to V4.1 Flash pending V4.1 Pro. However, DeepSeek’s later official API changelog says that, in response to user demand, it will continue providing V4 Pro API service after September 14, 2026, with billing unchanged. Developers that need reproducible behavior should therefore use the current documented model IDs and run regression tests rather than assuming a legacy route always maps to V4.1 Flash. 1523

A 552B MoE model that activates far less per token

V4.1 Flash is a Mixture-of-Experts (MoE) model with 552 billion backbone parameters. In an MoE system, only a selected subset of parameters is used for a given token instead of running the entire model for every step.

DeepSeek reports that V4.1 Flash activates 8B parameters during input processing (prefill) and 16B during output generation (decoding). It uses what the company calls a Causal Encoder–Decoder architecture. The distinction matters because long prompts and long generated responses impose different inference costs. 1735

Sparse activation does not make a model automatically cheap or fast in every deployment: actual throughput also depends on hardware, batching, quantization, serving software, context length, and request mix. But it is central to the model’s intended efficiency profile.

Why the 1M-token context window depends on KV-cache compression

V4.1 Flash supports contexts of up to one million tokens. 35

The hard operational problem with long contexts is the key-value (KV) cache—the stored attention state used while generating each additional token. As a prompt and response become longer, that memory can become a major serving constraint.

DeepSeek’s model card says its Causal Encoder–Decoder design projects the decoder’s global KV cache from the final encoder hidden states rather than deriving it separately from every decoder layer. The release also combines Compressed Sparse Attention 2 with FP4 KV caching. Together, DeepSeek reports a global KV-cache footprint of about 890 bytes per token, roughly one-quarter of DeepSeek V4 Flash’s corresponding footprint. 3539

That does not mean every one-million-token task will be economical or accurate. Context capacity is not the same as reliable retrieval, reasoning, or instruction-following across an entire prompt. Teams working with large codebases, document collections, or agent histories should test recall, latency, and cost on representative workloads.

Open weights and self-hosting

DeepSeek has published V4.1 Flash weights on Hugging Face under the MIT license. The model card describes the release as open weights, enabling organizations to download and deploy the model under that license’s terms. 35

This offers a different option from API-only models: teams can evaluate the model on their own infrastructure and control their serving stack. It does not eliminate deployment complexity—especially for a 552B-parameter backbone—but it gives users a path to self-hosting and deeper optimization.

Does V4.1 Flash beat V4 Pro?

DeepSeek says new pretraining methods and larger-scale reinforcement-learning post-training produced benchmark results ahead of flagship models, including V4 Pro. The company also cites tests by multiple parties that place V4.1 Flash ahead on performance, cost, speed, and total runtime. 17

Those are meaningful vendor-reported results, but they are not a universal verdict. Benchmark results can depend on task selection, prompting, tool configuration, scoring methodology, and the model version tested. Before migrating a production workflow, evaluate both models on the tasks that matter: difficult reasoning, coding acceptance rates, tool use, multimodal accuracy, reliability, safety controls, latency, and total cost.

Early OpenRouter usage: strong, but not a measure of all AI demand

V4.1 Flash arrived into a market where DeepSeek and other Chinese model providers already had substantial routed usage. OpenRouter ranks models by prompt and completion tokens processed through its own API. In its September 13 ranking, it listed DeepSeek V4.1 Flash at 4.94T tokens, marked as new; DeepSeek V4 Flash 0731 at 11.6T; V4 Flash 0423 at 4.36T; and Xiaomi MiMo-V2.5 at 7.77T. 57

Earlier reports also showed how rapidly usage rankings could change: one August snapshot put V4 Flash at 7.1T weekly tokens and reported that nine of the ten most-used models in that ranking were Chinese. 50

The important caveat is scope. OpenRouter token counts reflect traffic routed through OpenRouter—not worldwide AI usage, enterprise deployment, revenue, or model quality. They are useful evidence of developer-platform demand, but not proof that a model has won the broader market.

What V4.1 Flash means for developers

The strongest reason to evaluate V4.1 Flash is the combination of native multimodality, one-million-token context, open weights, and an architecture designed to reduce the memory cost of long-running inference. 1735

A sensible migration path is:

1. Move new Flash integrations to deepseek-flash.
2. Treat old Flash and Vision Exp IDs as compatibility routes, not stable model pins.
3. Benchmark V4.1 Flash against V4 Pro on your own prompts and success criteria.
4. Measure long-context retrieval and end-to-end latency separately from headline context-window capacity.
5. Use OpenRouter ranking data as a demand signal, not a substitute for production evaluation.

V4.1 Flash’s technical claims are substantial, particularly the 890-byte global KV-cache figure and sparse activation design. Its real significance will depend on whether those gains carry through to reliable, affordable performance on the workloads developers actually run.