01
Overview
DeepSeek V4.1 Flash is DeepSeek's new multimodal open-weight model released on September 10, 2026. It combines native image understanding, a one-million-token context window, a large Mixture-of-Experts architecture and a new production API identity, while DeepSeek is also preparing to route its older V4 Pro endpoint to the newer Flash model.
02
What is DeepSeek V4.1 Flash?
DeepSeek describes V4.1 Flash as a multimodal Mixture-of-Experts model with a 552-billion-parameter backbone. The model does not activate all of those parameters for every token: DeepSeek says it activates about 8 billion parameters per token during prefill and 16 billion during decoding.
The model accepts both images and text and produces text. Its official model card is published on Hugging Face under the MIT license, so the weights can be downloaded and used outside DeepSeek's hosted API subject to that license.
DeepSeek also offers V4.1 Flash through its API under the model name deepseek-flash.
03
Release date and availability
DeepSeek V4.1 Flash was released publicly on September 10, 2026. Reuters independently reported the launch the same day, while DeepSeek's official Hugging Face model page and API documentation provide the technical and access details.
There are two main ways to use it: DeepSeek's hosted API, or the public weights through supported local/self-hosted inference stacks. The official model page currently includes examples for Transformers, vLLM, SGLang and Docker-based serving.
04
1M-token context window
DeepSeek documents a context length of up to one million tokens for V4.1 Flash. On the hosted API, the current pricing documentation also lists a maximum output length of 384K tokens.
A large context window can be useful for long codebases, document collections, agent histories and other input-heavy workflows. It should not be interpreted as a guarantee that every application needs or benefits from filling the full context window; latency, memory use and cost still matter.
05
Native image understanding
V4.1 Flash is natively multimodal. DeepSeek's model card says a DeepSeek-ViT vision encoder and a two-layer MLP projector convert images into visual embeddings that are processed together with text embeddings.
The hosted deepseek-flash API supports DeepSeek's Vision feature. DeepSeek's pricing page currently lists vision support for V4.1 Flash while the older V4 Pro model does not support it.
This makes V4.1 Flash relevant not just for text and coding tasks but also for workflows involving screenshots, charts, documents, interfaces and other image-based inputs.
06
Causal Encoder-Decoder architecture and KV-cache compression
One of the most technically important changes is DeepSeek's Causal Encoder-Decoder architecture. The 40-layer Transformer is organized as a 20-layer causal encoder followed by a 20-layer decoder.
DeepSeek says this design, together with its sparse-attention and cache techniques, reduces the global KV-cache footprint to 890 bytes per token — roughly one quarter of DeepSeek V4 Flash. It also says persistent KV-cache storage is roughly one eighth of V4 Flash because of its SWA bounded-replay approach.
Those are first-party architectural claims from DeepSeek's model card, not independent measurements. They matter because KV-cache cost becomes increasingly important as context windows and agent workloads grow.
07
Controllable reasoning effort
V4.1 Flash supports a continuously controllable reasoning-effort setting from 1 to 100. DeepSeek presents this as a way to trade inference cost and latency against accuracy instead of relying only on a simple thinking/non-thinking switch.
The hosted API still supports both thinking and non-thinking modes. Developers should verify the exact request format in the current API documentation before deploying because model interfaces can change after launch.
08
DeepSeek V4.1 Flash API model name
The hosted model name is deepseek-flash.
DeepSeek says the legacy names deepseek-v4-flash and deepseek-v4-flash-vision-exp are still accepted, but those older models have been retired and requests using those names are now served by V4.1 Flash at Flash pricing.
This distinction matters for developers maintaining existing integrations: an old model identifier may continue working even though the underlying model has changed.
09
API pricing
DeepSeek currently prices V4.1 Flash differently during peak and off-peak periods. Its official pricing page lists the following rates per one million tokens:
- Cache-hit input: $0.003 off-peak and $0.006 peak.
- Cache-miss input: $0.15 off-peak and $0.30 peak.
- Output: $0.60 off-peak and $1.20 peak.
DeepSeek defines peak hours as 01:00–04:00 UTC and 06:00–10:00 UTC Monday through Friday, with other hours treated as off-peak. Prices are time-sensitive and should be rechecked on DeepSeek's official pricing page before budgeting a production workload.
10
What happens to DeepSeek V4 Pro?
DeepSeek says V4.1 Flash has surpassed V4 Pro in its own testing across performance, cost, speed and total time. That is a vendor claim rather than an independent benchmark conclusion.
More importantly for developers, DeepSeek has announced a concrete migration: from 12:00 Beijing Time on September 14, 2026, requests sent to deepseek-v4-pro will be routed to V4.1 Flash and billed at V4.1 Flash pricing until a future V4.1 Pro model is released.
If a production system depends on V4 Pro behavior, this change deserves explicit regression testing rather than assuming the replacement is identical.
11
Benchmarks
DeepSeek publishes extensive base-model and instruct-model results on the V4.1 Flash model card, including coding, agent and reasoning evaluations. These results are useful for understanding how DeepSeek positions the model, but they are primarily vendor-run evaluations.
For that reason, it is safer to treat the tables as reported evidence rather than declaring V4.1 Flash a universal benchmark winner. Independent reproduction will become more useful as the model reaches more inference providers and researchers.
12
Can you run DeepSeek V4.1 Flash locally?
Yes, the weights are public and the official Hugging Face page provides serving examples for vLLM and SGLang, along with Transformers and Docker-related paths. However, 'local' does not mean lightweight.
The model has a 552B-parameter backbone, and DeepSeek does not publish one universal minimum GPU or VRAM specification that applies to every precision, quantization, runtime and context length. Hardware guidance should therefore be treated cautiously. Community quantizations may reduce resource requirements, but they are separate artifacts from the official weights and should be labeled as such.
13
Who is V4.1 Flash for?
The strongest use cases are likely to include long-context coding, agentic workflows, multimodal document and interface analysis, high-throughput API workloads, and teams that want open weights for controlled deployment.
Its appeal is the combination of a large context window, native vision, public weights and comparatively low hosted API pricing. The tradeoff is operational complexity if you choose to self-host such a large model.
14
Important limitations
Several claims still need ongoing verification. DeepSeek's benchmark and efficiency comparisons are first-party results. Real-world quality will vary by workload. Local hardware requirements depend heavily on precision and deployment strategy. API pricing, routing and concurrency limits can change. Community GGUF or other quantized builds are not the same thing as the official checkpoint.
15
FAQ
Is DeepSeek V4.1 Flash open source?
The public model weights are released under the MIT license. 'Open-weight' is the more precise description for the model distribution itself, while the official repository/model card provides the implementation details needed to use it.
Does DeepSeek V4.1 Flash support images?
Yes. The official model card describes native image-and-text input, and DeepSeek's hosted API pricing/model page lists Vision support for deepseek-flash.
What is the context window?
DeepSeek documents up to one million tokens of context.
What API model name should I use?
Use deepseek-flash for the current V4.1 Flash hosted model. Legacy V4 Flash names are still accepted but route to V4.1 Flash.
Will DeepSeek V4 Pro disappear?
DeepSeek says that from September 14, 2026 at 12:00 Beijing Time, deepseek-v4-pro requests will route to V4.1 Flash until V4.1 Pro becomes available.
Is V4.1 Flash cheap to self-host?
That cannot be stated universally. The public weights give developers deployment freedom, but the model is very large. Actual infrastructure cost depends on precision, quantization, context length, throughput and hardware.
Sources
Primary and supporting sources
Facts were rechecked against the linked sources immediately before publication. Pricing, product availability and rollout status can change.