DeepSeek V4.1 Flash ships — and retires V4 Pro next Monday
DeepSeek's mid-morning release isn't just a new checkpoint. It's a new architecture, an open-weights drop, and a forced migration for anyone still calling its flagship — all in one announcement.
DeepSeek released V4.1 Flash on Wednesday, the first model in a new architecture family and the first with native visual understanding. The headline number is 552 billion parameters, but the point is how few of them run: a new causal encoder–decoder design activates just 8 billion parameters while reading input and 16 billion while writing output. The technical report adds a separate 196 billion parameters of n-gram "engram" memory on top of the 552B backbone — the sort of lookup table that can sit on disk rather than in GPU memory, which is why local-runner forums spent the morning arguing about whether this fits on 192 GB.
The cost story is the part that lands. DeepSeek claims the new KV cache needs a quarter of the HBM and an eighth of the SSD storage of the previous generation, and says cache-hit charges — the dominant line item in most agent bills — fall accordingly. Weights and the technical report are on Hugging Face, and peak/off-peak pricing continues with off-peak at half of peak. V4 Pro is being retired: after noon Beijing time on September 14, calls to it route to V4.1 Flash and bill at Flash prices until V4.1 Pro arrives. That is a lab saying its cheap model beats its expensive one, and acting on it within four days.
Treat the benchmark claim with the usual salt. DeepSeek says V4.1 Flash surpasses V4 Pro and several other flagships, but the comparison tables it published don't include the newest closed models, and one independent reading found it still trailing on several agentic coding evals while actually losing to Pro on dashboard design. The honest summary: a very cheap model with near-flagship agent scores, not a new frontier. It also lands one day after DeepSeek reversed its August price increases while preparing a Shanghai listing — we covered that in DeepSeek preps a STAR Market IPO — and cuts prices 60% the same day.
Princeton's Mengdi Wang used the Bund Summit stage to argue that AI has not yet produced a genuine basic-science discovery — and that scaling won't fix it. Her diagnosis is that models are trained to find the most likely answer, while real discoveries live in the long tail; in "simulated life" experiments with ChatGPT and Claude, the models consistently over-weighted the common case and under-weighted rare ones. Math and code progressed fast because both have cheap verifiers — a compiler, a unit test, a formal proof checker — that let a model try, fail, and correct. Wet-lab science mostly doesn't.
Her prescription is infrastructure, not parameters: her group has automated a graphene experiment workflow that used to take researchers months and exposed the equipment as an API others can call and reproduce. "Ideas are cheap" was her line — what's missing is a world that can be checked. It's the sharpest version yet of the case that the next real gain comes from making experiments verifiable, not from another order of magnitude of compute.
China has started work on its first two agent-security standards, launched at the Bund Summit by the Cyber Security Association of China with Ant Group and a consortium of carriers, platforms, and research institutes. One covers agent identity and authorization — who the agent is, who it acts for, what it may touch, with permissions that activate per task and are revoked when the task ends. The other covers runtime: monitoring whether an agent's behavior has drifted from the user's original intent, and what to do when it has.
The framing is the interesting part. Ant's security chief described agents as a new kind of subject in digital systems, not another object to be firewalled — one with an identity and permissions that can still misbehave through hallucination or prompt injection. A legitimate identity, in other words, no longer implies a legitimate action. These are voluntary group standards, not regulation, but they're the scaffolding Chinese regulators tend to reach for later.
What to watch: whether anyone reproduces DeepSeek's agent scores independently before the V4 Pro cutoff on Sunday, and whether a 552B model that runs like a 16B one becomes the template every other lab copies.
Is retiring your own flagship four days after launching its replacement a sign of confidence — or of a roadmap that got away from you? Tell us in the comments.
