DeepSeek has given developers four days to test a model migration they cannot opt out of. From 12:00 Beijing time on September 14, every API request using deepseek-v4-pro will be served by the new V4.1 Flash and billed at the Flash rate, according to the company's current model and pricing page. Today's open-weight release therefore comes with an operational deadline for any application pinned to the Pro identifier. Its behavior may change even if its code and model string do not.
That deadline deserves more attention than the usual launch-day benchmark contest. DeepSeek describes V4.1 Flash as a 552-billion-parameter multimodal mixture-of-experts model, yet it activates 8 billion parameters per token during input processing and 16 billion during generation. The official model card says this split is meant to cut the cost of long, input-heavy agent tasks. It also publishes MIT-licensed weights, a reference runtime and the encoding code needed to reproduce the prompt format.
The name "Flash" can make the release sound like a smaller edition of V4 Pro. DeepSeek's architecture says otherwise. V4.1 Flash is a public member of a new 40-layer causal encoder-decoder family, and its 552 billion backbone parameters sit between the earlier 284-billion-parameter V4 Flash and 1.6-trillion-parameter V4 Pro. DeepSeek has paired that architecture change with the decision to retire both older Flash models now and reroute V4 Pro traffic four days later, as its model card and API table document.
One model name now matters
The stable API name is now deepseek-flash. DeepSeek says the legacy identifiers deepseek-v4-flash and deepseek-v4-flash-vision-exp are still accepted, but both already resolve to V4.1 Flash. The deepseek-v4-pro identifier will join them on September 14 and remain an alias until a future V4.1 Pro arrives; the company gives no release date for that model on its pricing page. An alias preserves request compatibility, but it cannot promise identical answers, latency or tool behavior.
The new endpoint accepts text and images, supports tool calls and exposes both OpenAI Responses and Anthropic-compatible interfaces. DeepSeek lists a one-million-token context window and a maximum output of 384,000 tokens for V4.1 Flash. V4 Pro has the same stated context, output and interface support, except for vision. Those matching API checkboxes make the switch easy at the transport layer. They do not establish behavioral equivalence between two models with different architectures and post-training, a distinction visible in DeepSeek's side-by-side API specifications.
Price changes are large enough to explain the consolidation. At off-peak rates, one million uncached input tokens cost $0.15 on V4.1 Flash and $0.66 on V4 Pro; one million output tokens cost $0.60 versus $1.98. Peak prices are twice those figures. DeepSeek also lists a concurrency limit of 2,500 for Flash and 500 for Pro on the same rate card. These are posted prices, rather than a measured total cost for an agent job, where token use and retries can erase part of the per-token saving.
The model spends compute differently on input and output
V4.1 Flash divides its Transformer into a 20-layer causal encoder followed by a 20-layer decoder. During prefill, when the model processes the prompt, DeepSeek says only 8 billion parameters are active per token. Decode activates 16 billion as the model produces its answer. The decoder's global key-value cache is projected from the encoder's final hidden states instead of being rebuilt from every decoder layer, according to the model card.
That arrangement targets the awkward economics of coding agents. A repository, tool history and prior reasoning can make the input much larger than the final patch or explanation. DeepSeek reports a persistent global KV-cache footprint of 890 bytes per token, about one quarter of V4 Flash's figure, in its architecture description. At one million tokens, the published rate implies roughly 890 MB before serving overhead and any other model state. The number is a design claim from DeepSeek; independent serving tests will need to show how much memory a real deployment saves.
Compressed Sparse Attention 2 supplies part of that reduction. Layers use one of three fixed attention modes: Full, Reindex or Reuse. Later decoder layers search within a candidate pool built by the first Full layer, while the main KV cache uses FP4 values with a scale for each 16 channels. DeepSeek also adds a 196-billion-parameter conditional-memory component called Engram, according to the released architecture notes. Those parameters are accessed through token-based lookup, so the headline parameter count says little by itself about work per token.
The training disclosure is substantial but still comes from the developer. DeepSeek says in the model card that it trained V4.1 Flash from scratch on 45 trillion multimodal tokens, trained sparse attention at 64,000 tokens and extended context to one million tokens after 34 trillion training tokens. Post-training used supervised fine-tuning, reinforcement learning and on-policy distillation, with an integer reasoning-effort control from 1 to 100. The company has not published the training corpus in the release materials.
DeepSeek's own tables argue for regression testing
DeepSeek says V4.1 Flash has surpassed V4 Pro in performance, cost, speed and total time. Its own benchmark table is more mixed. At maximum reasoning effort, V4.1 Flash scores 90.6 on Terminal-Bench 2.1 against V4 Pro's 87.9, and resolves 74.2 percent of DeepSWE v1.1 tasks against 62.7 percent. It also leads the two older DeepSeek models on CyberGym and AutomationBench in the published evaluation.
Other rows go the other way. On the base-model table, V4 Pro scores 55.2 on SimpleQA-Verified while V4.1 Flash scores 42.3. V4 Pro also leads on LongBench-V2, 51.5 to 45.2, and on MATH, 64.5 to 61.1. In the instruction-model comparison, V4.1 Flash records 36.8 on Humanity's Last Exam, or 39.1 on the text-only subset, below V4 Pro's reported 42.7 on that subset. A single claim of overall superiority hides differences that may matter more than the aggregate for a particular product.
The agent scaffold changes the result too. DeepSeek reports a 74.2 DeepSWE score with mini-SWE, 72.6 with its DSH Minimal scaffold and 65.6 with Codex. On Terminal-Bench 2.1, the same model ranges from 84.1 with Codex to 90.6 with DSH Minimal. The evaluation notes specify eight samples per DeepSWE task, three per Terminal-Bench task, up to 500 agent steps and a one-million-token context. Those settings help interpretation, but they are far from a drop-in prediction for an existing agent loop.
A 510 GB reference release
The repository is genuinely open-weight rather than an API announcement with a download promised later. Its Hugging Face record lists 48 weight shards, about 510 GB of stored files and an MIT license. The uploaded tensors use a mixture of FP8 and FP4 formats, while the configuration describes 384 routed experts per mixture-of-experts layer and six selected experts per token.
Running those weights is a separate engineering project. DeepSeek labels its included implementation a readable reference rather than a production server. The inference guide converts the checkpoint across eight tensor-parallel ranks, and its sample launch runs torchrun with eight local processes. The repository also lacks a Jinja chat template. Instead, it ships a Python encoding implementation with test vectors and points production users to deepseek-recipe, a separate set of Rust libraries with Python bindings for Messages, Chat Completions and Responses requests.
That prompt-format detail is easy to miss during an API migration. DeepSeek says the released prompt encoder handles tool calls, reasoning-effort values, mid-conversation system messages and interleaved images. A self-hosted server that treats V4.1 Flash like an older Transformers chat model may tokenize the conversation incorrectly even after the 510 GB checkpoint loads. The release provides encoding code and test cases for implementers to compare against.
What to test before the alias flips
Teams using deepseek-v4-pro have a narrow but useful window. The first check is to send a representative evaluation set to both the Pro identifier and deepseek-flash before September 14, holding the system prompt, reasoning effort and tool schema constant. Structured-output validity, tool selection, token consumption and completion time are more informative for that decision than DeepSeek's cross-model averages. The API table confirms that both identifiers currently remain callable.
Self-hosters have a different watch list. Serving projects need native support for the deepseek_v41 configuration, its causal encoder-decoder layout, CSA2 attention, FP4 expert weights and the released encoder. DeepSeek's reference inference guide can check tensor shapes and kernel plumbing, but its self-test uses uninitialized weights and does not test numerical correctness. A successful import is therefore only the start of validation.
The next evidence should come from independent reproductions of the agent scores, memory measurements from production inference engines and migration reports after the September 14 reroute. DeepSeek has left the timing of V4.1 Pro open on its current model page. Until that model appears, the most revealing result will be whether applications tuned for V4 Pro keep their quality when the identifier begins serving a cheaper model with a very different compute path.