DeepSeek V4.1 Flash is not a routine refresh of V4 Flash. DeepSeek has changed the model's architecture, expanded its native multimodal support, cut the storage required for long-context inference, and introduced a new API price schedule. The company is also positioning V4.1 Flash as the temporary destination for V4 Pro traffic while it prepares a future V4.1 Pro.
The result is a model aimed squarely at agent workloads: long prompts, repeated tool calls, large repositories, images mixed with text, and enough cached context to make memory cost matter. Its headline specifications are a 552-billion-parameter Mixture-of-Experts backbone, 8 billion active parameters during input processing, 16 billion during generation, a one-million-token context window, and native image understanding.
Those numbers are useful, but they do not answer the practical question by themselves. The more interesting change is how DeepSeek separates input processing from output generation and compresses the KV cache that keeps long conversations and agent traces available during inference.
DeepSeek V4.1 Flash is available through the DeepSeek API as
deepseek-flash. Teams already using the retired V4 Flash aliases should check routing and billing behavior before comparing old and new runs.
DeepSeek V4.1 Flash at a glance
| Detail | DeepSeek V4.1 Flash |
|---|---|
| API model name | deepseek-flash |
| Model type | Multimodal Mixture-of-Experts (MoE) |
| Backbone parameters | 552B |
| Active parameters | 8B during prefill; 16B during decode |
| Architecture | 40-layer Causal Encoder-Decoder: 20-layer causal encoder plus 20-layer decoder |
| Context length | 1M tokens |
| Maximum API output | 384K tokens |
| Inputs | Text and images |
| Outputs | Autoregressive text |
| Reasoning control | Continuous reasoning_effort from 1 to 100 |
| Open weights | Yes, under the MIT License |
| OpenAI-compatible base URL | https://api.deepseek.com |
| Anthropic-compatible base URL | https://api.deepseek.com/anthropic |
DeepSeek says the model supports both thinking and non-thinking modes. JSON output, tool calls, the Responses API, the Anthropic API, chat prefix completion, and FIM completion are documented, although FIM is limited to non-thinking mode.
DeepSeek model-card figure on KV cache compression. This is vendor-published evidence and does not independently establish production cost or latency.
Why the Causal Encoder-Decoder architecture matters
Most decoder-only language models repeatedly carry KV states across many layers while processing a growing context. That becomes expensive when an agent keeps a large codebase, tool history, screenshots, and intermediate results in one session.
DeepSeek V4.1 Flash uses a Causal Encoder-Decoder, or CED, design. Its 40 Transformer layers are divided into a 20-layer causal encoder and a 20-layer decoder. Instead of deriving the decoder's global KV cache independently from every decoder layer, the model projects it from the encoder's final hidden states.
This creates an intentional asymmetry. During prefill, when the model reads the prompt and existing context, it activates 8B parameters per token. During decode, when it produces the response, activation rises to 16B. That distinction fits agent workloads where the input may be much larger than the output.
DeepSeek also introduces Compressed Sparse Attention 2. Attention layers use one of three static modes: Full, Reindex, or Reuse. The design shares KV data and sparse-attention indices across layers instead of rebuilding everything repeatedly. In the decoder, a hierarchical sparse indexer narrows later searches to a candidate pool created by the first full-attention layer.
The model card reports a global KV cache footprint of 890 bytes per token. DeepSeek says that is roughly one quarter of V4 Flash and about 1/437 of DeepSeek V1. Its SWA Bounded Replay mechanism also avoids persisting the full sliding-window-attention KV state to SSD, reducing persistent cache storage to about one eighth of the V4 Flash requirement.
These are architecture and storage claims from DeepSeek's own model card. They make V4.1 Flash interesting for long-running agents, but they do not guarantee lower end-to-end cost in every deployment. Prompt reuse, cache hit rate, output length, inference provider, concurrency, and tool latency still determine the bill users actually see.
Native vision and long-context agent work
V4.1 Flash processes images and text together. DeepSeek describes a vision encoder trained from scratch with 2D-RoPE and 3x3 pixel-unshuffle downsampling, followed by a two-layer projector that converts images into embeddings for the language model.
That native multimodal path matters for agents that need to read charts, screenshots, documents, interfaces, or visual test results without handing the image to a separate model. It also changes how the new model should be compared with the original V4 Flash, which did not represent the same multimodal product boundary.
The one-million-token context window gives developers room for large repositories and long trajectories, but capacity is not the same as reliable recall. Long-context testing should measure whether the model finds the right evidence, follows instructions buried in the trace, and avoids carrying stale assumptions into later steps. A large advertised window only establishes the accepted input limit.
DeepSeek V4.1 Flash API pricing
DeepSeek prices the API per one million tokens and separates input into cache hits and cache misses. It also uses peak and off-peak rates. The table below replaces the Chinese pricing graphic in the source announcement with text taken from the official DeepSeek API pricing page.
| Token category | Off-peak price per 1M tokens | Peak price per 1M tokens |
|---|---|---|
| Input, cache hit | $0.003 | $0.006 |
| Input, cache miss | $0.15 | $0.30 |
| Output | $0.60 | $1.20 |
Prices above were verified on September 10, 2026. DeepSeek defines peak hours as 01:00-04:00 UTC and 06:00-10:00 UTC, Monday through Friday. All other hours use the off-peak rate. The official page says off-peak prices are half of peak prices.
The gap between a cache hit and a cache miss is substantial. At peak pricing, one million uncached input tokens cost $0.30, while the same number of cached input tokens cost $0.006. That makes stable prompt prefixes, reusable repository context, and repeated agent instructions economically important. It does not mean every repeated token will automatically receive the cache-hit rate; teams need to measure actual billing data rather than assume a theoretical hit ratio.
For comparison, the official pricing page still lists deepseek-v4-pro at the following rates before its announced routing change:
| V4 Pro token category | Off-peak price per 1M tokens | Peak price per 1M tokens |
|---|---|---|
| Input, cache hit | $0.022 | $0.044 |
| Input, cache miss | $0.66 | $1.32 |
| Output | $1.98 | $3.96 |
DeepSeek says requests to deepseek-v4-pro will be routed to V4.1 Flash from 12:00 Beijing time on September 14, 2026 until V4.1 Pro is released. Those routed requests will be billed at the V4.1 Flash rate. This is a scheduled migration, not evidence that V4.1 Pro is already available.
The documented concurrency limits are 2,500 for deepseek-flash and 500 for deepseek-v4-pro. Rate limits and service isolation can affect production throughput independently of the per-token price.
API model names and migration behavior
New integrations should use:
DeepSeek says the older names deepseek-v4-flash and deepseek-v4-flash-vision-exp remain accepted for compatibility, but the original models behind those names have been retired. Requests using either alias are served by DeepSeek V4.1 Flash and billed at the Flash price.
That compatibility layer reduces migration work, but it can also make benchmark records confusing. A log that says deepseek-v4-flash after the routing change may no longer represent the model originally tested under that name. Store the request date, returned model metadata where available, reasoning settings, API endpoint, and pricing schedule with every evaluation.
Developers should also recheck prompts. V4.1 Flash supports continuously adjustable reasoning effort, and DeepSeek's published instruct evaluations use reasoning_effort=100, temperature=1.0, and top_p=0.95. A cheaper or faster production setting may not reproduce those scores.
DeepSeek model-card benchmark comparison. Results are vendor-published and depend on the stated harnesses, settings, and evaluation tasks.
What the published benchmarks do and do not show
DeepSeek's model card reports strong results on several coding, security, automation, and visual-agent tests. For example, it lists 90.6 on Terminal-Bench 2.1, 74.2 resolved on DeepSWE v1.1, 88.1 on CyberGym, and 54.8 on AutomationBench. It also reports weaker relative results on some other tests, including Terminal-Bench 3.0 and 4.0, where Opus 5.0 leads the table.
This mixed pattern is more informative than a single "beats frontier models" headline. V4.1 Flash appears particularly competitive on some agent tasks, but no one score settles model choice. The published results use different harnesses for different benchmarks, including DeepSeek Harness, mini-SWE-agent, Claude Code, and official benchmark scaffolds. Sample counts, network access, maximum steps, context limits, and reasoning effort also vary by setup.
All of these figures come from DeepSeek's release materials rather than an independent Atoms evaluation. They are useful for deciding what to test, not for promising the same success rate in a production workflow.
A practical evaluation should keep the task and harness fixed, then compare:
- accepted task completion rather than raw token throughput;
- total input, cached input, and output cost per successful run;
- wall-clock time, including tools and retries;
- repository or document scale;
- failure recovery and instruction retention;
- output quality under the reasoning setting you can afford.
Open weights do not make deployment lightweight
DeepSeek released the V4.1 Flash weights and model repository under the MIT License. The repository includes prompt encoding guidance, inference notes, and instructions for reproducing selected DeepSWE results.
The model is still large. The announcement asks organizations interested in large-scale deployment to contact DeepSeek and cites an example resource profile of roughly 2,000 GPU cards plus a storage cluster. Most teams will evaluate the API or a managed inference provider before considering self-hosting.
Open weights are still meaningful. They allow infrastructure vendors and research groups to inspect the model artifacts, adapt inference engines, reproduce parts of the evaluation, and study the architecture without waiting for a closed API. They do not remove the engineering cost of serving a 552B-parameter multimodal MoE model.
Who should test DeepSeek V4.1 Flash
The model is most relevant when input dominates the workload: coding agents reading large repositories, research agents carrying many source documents, multimodal agents interpreting screenshots, or repeated workflows with a stable cached prefix.
It is less obvious as a default when requests are short, output quality matters more than input cost, or a team cannot control when peak rates apply. The announced prices are low, but the cheapest token is not always the cheapest completed task. A model that needs more retries can erase an apparent price advantage.
For an initial trial, use a small set of real tasks and record both quality and billing. Keep the same harness and tool permissions across models. Test at more than one reasoning-effort setting. Include at least one long-context task and one multimodal task if those capabilities are part of the buying decision.
You can also compare the existing DeepSeek V4 Flash model page with other models in the Atoms model catalog. Because DeepSeek now routes legacy Flash names to V4.1 Flash, verify the actual runtime model before treating an older label as a separate model.
Bottom line
DeepSeek V4.1 Flash is built around a concrete systems problem: agent workloads are input-heavy, context grows quickly, and KV cache storage can become a major cost. Its asymmetric Causal Encoder-Decoder design, sparse attention, and compressed cache attack that problem directly.
The API offer is equally aggressive. Off-peak uncached input costs $0.15 per million tokens, cached input costs $0.003, and output costs $0.60. Peak rates are double. Those prices, the one-million-token context window, native vision, and open weights make V4.1 Flash worth testing for coding and multimodal agents.
The right conclusion is still workload-specific. DeepSeek's benchmarks are vendor-published, legacy model names now route to new infrastructure, and the lowest token price does not measure retries or accepted output. Evaluate the model under your own harness, preserve the exact API and reasoning settings, and compare cost per successful task.
Use Atoms to build what you want easily
Atoms helps turn your idea into an editable product workflow. You can start from a natural-language brief, visualize the layout and content, and iterate on the result before treating it as an MVP or internal tool. It is not a replacement for production engineering review, but it is useful when you need to test a visual concept in a working interface.
Frequently asked questions
What is the DeepSeek V4.1 Flash API model name?
Use deepseek-flash. DeepSeek says the legacy names deepseek-v4-flash and deepseek-v4-flash-vision-exp are temporarily accepted but route to V4.1 Flash.
How much does DeepSeek V4.1 Flash cost?
At off-peak rates, one million cached input tokens cost $0.003, uncached input costs $0.15, and output costs $0.60. Peak rates are $0.006, $0.30, and $1.20 respectively. These prices were verified on September 10, 2026 and may change.
What is the context window?
The official API documentation lists a one-million-token context window and a maximum output of 384K tokens.
Does DeepSeek V4.1 Flash support images?
Yes. DeepSeek describes it as a native multimodal model that processes image and text inputs and generates text.
Is DeepSeek V4.1 Flash open source?
DeepSeek publishes the model weights and repository under the MIT License. Running the 552B-parameter model still requires substantial compute and storage infrastructure.
Is DeepSeek V4 Pro being replaced?
DeepSeek says API requests to deepseek-v4-pro will route to V4.1 Flash from September 14, 2026 at 12:00 Beijing time until a future V4.1 Pro is released. That temporary routing does not mean V4.1 Pro is already available.
