Labs earlier regarded a model drop as an occasion which reset the scoreboard.
For a stretch of 2024 and early 2025 that was not only advertising speak. A launch may move chip stocks, deplete a weekend of analyst notes, and drive every different group to clarify why its value curve nonetheless made sense. That environment has lessened. Updates nonetheless arrive on schedule. They're denser, extra particular about cache bytes and routing tables, and they arrive in a market that has already priced within the idea that anyone in Hangzhou can practice a critical system with out a Western campus and a Western invoice.
DeepSeek was the lab that made that concept public.
In January 2025 its V3 and R1 launches did what a protracted run of Chinese language fashions had not: it's models shocked Silicon Valley and sat above the mighty ChatGPT within the U.S. retailer.
Unfortunately, the shock didn't repeat.
But the development continues.
But not as a second reset.
DeepSeek announced 'V4.1-Flash,' which is essentially the smallest mannequin in a brand new structure household. It is a 552 billion parameter mixture-of-experts system with native picture understanding and a million-token context.
The design the corporate needs recognized is just not but one other bounce in uncooked parameter depend.
According to its paper, it's an asymmetrical Causal Encoder-Decoder stack: 40 layers break up into a twenty-layer causal encoder and a twenty-layer decoder, with solely about 8 billion parameters energetic whereas the mannequin reads a immediate and 16 billion whereas it writes.
Prefill, the costly a part of lengthy agent traces, is meant to run at roughly half the earlier compute.
Compressed Sparse Attention 2, FP4 key-value storage, and a replay trick for sliding-window states reduce the dwell KV cache to roughly a quarter of the earlier Flash era's HBM use and the everlasting cache to roughly an eighth of the SSD footprint.
Cache hits are a big share of agent payments. Shrinking the cache is the way in which the lab argues it might probably hold serving long classes with out asking customers to pay for final yr's reminiscence tax.
On the API the mannequin is deepseek-flash.
The older V4-Flash and the short-lived visible experiment now route into it.
DeepSeek mentioned third-party exams put V4.1-Flash forward of V4-Pro on high quality, value, velocity, and complete runtime, and that Pro visitors would shift to Flash at Flash charges from 14 September untill a V4.1-Pro exists.
Companions listed in the identical thread, Work Buddy (together with Code Buddy) and Open Code, already level on the new weights. Official pricing retains a peak and off-peak break up, with off-peak at half of peak, and the corporate minimize Flash charges when the structure change made serving cheaper.
Weights and a technical report went up for anybody who desires to host the stack themselves. The final put up within the thread requested folks planning 2,000-GPU deployments with a storage cluster to get in contact.
The numbers on the card are combined in the best way present benches sometimes are.
At most reasoning effort the lab stories 90.6 on Terminal-Bench 2.1, a hair above the Claude Opus 5 and GPT-5.6 Sol figures it printed subsequent to it, 90.9 on GPQA Diamond, a 3471 Codeforces ranking, and 74.2 on DeepSWE.
Different rows, together with later Terminal-Bench variations, sit behind the identical closed fashions.
None of these comparisons have been independently audited within the announcement itself.
A day earlier than the thread, U.S. businesses named DeepSeek amongst Chinese language labs they mentioned had distilled American fashions at industrial scale. That cost and this launch now sit in the identical information cycle whether or not or not they belong collectively.
What the thread does not appear to be is January 2025. There isn't a declare of a brand new intelligence class. There's only a declare that a smaller, much less expensive serving shape can exchange the corporate's personal flagship for many site visitors.
Builders who had pinned evaluations and prompts to V4-Pro protested an abrupt swap; the cutover window was stretched after the primary discover, and changelog textual content later wobbled on whether or not Pro would stay a dwell endpoint.
That form of friction is routine product work. It's additionally the precise reverse of the myth the primary DeepSeek shock created, which was that a lean lab may preserve leaping whereas everyone else was nonetheless budgeting GPUs.
The sincere learn is narrower than both myth. DeepSeek nonetheless publishes structure that different teams copy notes from. It nonetheless costs inference as if reminiscence and prefill have been the actual constraints, which they're for brokers that reread lengthy traces.
The West was disrupted as soon as by the proof that frontier-adjacent fashions may be educated and bought at a unique value curve. The present tense is a lab squeezing that curve, retiring its personal high SKU into the environment friendly one, and transferring on the velocity of cache layouts slightly than on the velocity of a panic on Nasdaq.
