What Happened
On September 10, DeepSeek pushed a notice to its API customers that read less like a product launch and more like a eulogy. The Hangzhou company was releasing V4.1-Flash, the smallest member of a new architecture family, and it was retiring V4-Pro — the flagship it had spent the year selling as its best work. Four days later, at noon Beijing time on September 14, every request still pointed at deepseek-v4-pro began quietly resolving to the cheaper model instead, billed at Flash rates. DeepSeek's own wording on the matter was blunt: "We're phasing out V4-Pro."
The replacement is not a trimmed-down version of the old flagship. V4.1-Flash runs a 552-billion-parameter mixture-of-experts backbone on what DeepSeek calls a causal encoder–decoder design, and the split matters more than the headline number. Reading your input activates roughly 8 billion parameters. Writing the response activates about 16 billion. Independent trackers put the full parameter count closer to 763 billion once conditional-memory weights are included, trained across some 45 trillion tokens. The model accepts images natively, holds a million tokens of context, and ships under an MIT license with weights on Hugging Face — no gated access, no research-only clause.
Pricing moved the same day the model did. On DeepSeek's own API, off-peak rates land at roughly one yuan per million cache-miss input tokens and four yuan per million output tokens, with cache hits at two fen. Peak hours double that. In dollar terms the platform lists about $0.30 per million input and $1.20 per million output, while third-party hosts including Fireworks, DeepInfra and Novita have gone lower still. The older V4-Flash and V4-Flash-Vision-Exp endpoints are gone, temporarily redirected for compatibility.
Why It Matters
Most labs handle a generational jump by stacking a new premium tier on top of the old one. You get a Pro, then an Ultra, then something with a longer name and a higher price, and the previous flagship slides down into the budget slot. DeepSeek did the opposite. It looked at a model that beat its flagship on performance, cost, speed and total completion time, and then deleted the flagship rather than repositioning it. That is a pricing decision disguised as a product decision, and it only works if the economics underneath are real.
They appear to be. The number DeepSeek chose to lead with in its technical write-up was not a benchmark score but a memory figure: compared with the previous generation, V4.1-Flash needs a quarter of the HBM and an eighth of the SSD storage to hold the same key-value cache. For anyone running agents — long chains of tool calls where the same prompt prefix gets re-read hundreds of times — cached input is often the single largest line on the bill. Shrinking the cache does not just make inference faster. It changes what a company can afford to charge.
The asymmetry in the architecture points the same direction. A model that spends 8 billion active parameters reading and 16 billion writing is optimized for a world where inputs are enormous and outputs are short: document analysis, codebase navigation, retrieval-heavy agent loops. That is precisely the workload profile that has been eating margins across the industry, and it is the one DeepSeek has now priced most aggressively against.
There is a competitive dimension too. Two of the models most frequently compared against V4.1-Flash are closed and considerably more expensive. A million-token, image-capable, MIT-licensed model that can be downloaded and self-hosted puts a hard ceiling on what anyone can charge for the same capability class — not because enterprises will all switch, but because procurement teams now have a number to point at.
Reaction
The open-weights release landed first with the people who actually run inference. Within days the weights were mirrored, quantized and being wired into serving stacks, and DeepSeek said it would work with the open-source community on inference support rather than leaving that to chance. Two launch partners, WorkBuddy and OpenCode, shipped support immediately — a modest list, but a signal that the coding-agent crowd was the intended audience from the start.
The V4-Pro retirement drew a cooler response. Teams that had tuned prompts, evaluated outputs and signed off on a specific flagship model found it silently swapped underneath them on a four-day notice. DeepSeek's framing — that multiple parties had tested the replacement as faster, cheaper and better — is the kind of claim that reads well in a blog post and poorly in a change-management meeting. The company did soften the landing by billing the redirected traffic at the lower Flash rate, which means most affected customers woke up on September 14 paying less than they had the day before.
What's Next
The obvious missing piece is V4.1-Pro. DeepSeek has said the deepseek-v4-pro redirect holds only until that model arrives, which is an unusually specific promise for a company that rarely pre-announces. Read alongside the description of V4.1-Flash as "the smallest model in our new architecture family," the roadmap draws itself: the encoder–decoder design is meant to scale, and Flash is the proof of concept shipped first because it is cheapest to serve.
The financial backdrop is harder to ignore. Days before the launch, Reuters reported that DeepSeek had engaged CITIC Securities to prepare a listing on Shanghai's STAR Market, with the aim of starting the process this year. A pre-IPO round has been discussed at a valuation near 500 billion yuan, roughly $75 billion, though neither the raise size nor the listing valuation has been settled. A company weeks away from opening its books to regulators has reason to show a clean, consolidated product line and a cost curve bending the right way. Retiring an expensive flagship in favour of a model that is cheaper to serve does both.
DeepSeek also slipped an invitation into the launch post aimed at anyone planning a deployment of 2,000 GPUs or more. That is not a note to hobbyists. It suggests the company intends to sell infrastructure-scale partnerships alongside API tokens, which would be a meaningful widening of a business that has so far been mostly tokens and goodwill.
Closing Thoughts
What makes this launch worth watching is not the benchmark table. It is the willingness to kill a flagship. Frontier labs have spent three years training the market to believe that better means bigger and bigger means pricier, and the entire tiering structure of the industry rests on that association holding. DeepSeek just released a model it describes as smaller, gave it away under an MIT license, and used it to replace the expensive thing at the top of its own menu.
If the efficiency claims hold up under sustained load — and serving a million-token context to a large user base is where architectural elegance usually meets its limits — then the interesting question is not whether DeepSeek can keep undercutting everyone. It is how long the premium tier stays premium anywhere. A closed model priced at ten times an open one has to be ten times better at something the customer actually does, and that argument gets harder to make every quarter.
The counterweight is that open weights and a low sticker price are easier to sustain before an IPO than after one. Public markets tend to ask when the pricing stops being strategic and starts being a business. For now, though, a company in Hangzhou has made the cheapest option in its own catalogue the best one, and left everyone else to explain why theirs is not.
Sources: DeepSeek's launch announcement, TechNode, LLM Stats, and the model weights on Hugging Face.
한글 요약
딥시크가 9월 10일 새 모델 V4.1-Flash를 공개하면서, 동시에 자사 최상위 모델이던 V4-Pro를 단종하겠다고 알렸습니다. 9월 14일 베이징 시간 정오부터 V4-Pro로 들어오던 API 요청은 전부 V4.1-Flash로 넘어가고, 요금도 더 저렴한 Flash 기준으로 매겨집니다. 새 모델은 552B 규모의 MoE 구조에 입력 처리 8B·출력 생성 16B만 활성화하는 비대칭 설계를 썼고, 100만 토큰 컨텍스트와 이미지 입력을 지원합니다. 가중치는 MIT 라이선스로 허깅페이스에 공개됐습니다.
회사가 가장 앞세운 수치는 벤치마크 점수가 아니라 메모리였습니다. 이전 세대와 비교해 KV 캐시가 쓰는 HBM은 4분의 1, SSD는 8분의 1로 줄었다는 것입니다. 같은 프롬프트 앞부분을 수백 번 다시 읽는 에이전트 작업에서 캐시 비용은 청구서의 가장 큰 항목인 경우가 많습니다. 캐시를 줄였다는 건 속도만이 아니라 가격을 어디까지 내릴 수 있느냐의 문제이기도 합니다. 비공개 경쟁 모델들과 비교하면, 내려받아 직접 돌릴 수 있는 이 모델의 존재 자체가 같은 성능대 가격의 상한선 역할을 하게 됩니다.
배경에는 상장 준비가 있습니다. 로이터는 딥시크가 상하이 커촹반(STAR 마켓) 상장을 위해 CITIC증권을 선임했고 연내 절차 착수를 목표로 한다고 전했습니다. 상장 전 투자 라운드에서 거론되는 기업가치는 5,000억 위안, 약 750억 달러 수준입니다. 비싼 최상위 모델을 정리하고 서빙 비용이 싼 쪽으로 제품군을 단순화하는 선택은, 곧 장부를 공개해야 하는 회사 입장에서 보면 제품 결정인 동시에 재무 결정이기도 합니다. 다음 순서는 예고된 V4.1-Pro이고, 이번 Flash는 그 구조가 확장 가능하다는 것을 먼저 증명해 보인 셈입니다.