What even happened to deepseek

DeepSeek arrived in January 2025 like a stone dropped into still water. R1 was the impact; everything since has been the rings spreading outward. By late March, the forums had moved from excitement to a kind of puzzled silence — "what even happened to DeepSeek?" was a genuine question, asked by people who had expected the follow-through and found only quiet. The answer, it turned out, was that they were building. On April 10th, early V4 benchmarks began circulating on Singularity, scoring high enough to earn attention without yet earning certainty. Then April 24th arrived, and the picture snapped into focus all at once: DeepSeek V4 landed on Hacker News with a score of 8.2, accompanied immediately by benchmark breakdowns and, almost comically, a disclosure that the model carries a 384,000-token max output capability — a number that made several practitioners stop mid-sentence and reread it. The architecture is built around a mechanism called mHC; someone ran the internals on 8x H100s and published what they found, and the community spent the following days arguing about what it actually implies. Not every signal was celebratory — V4 Pro underwhelmed on the Arena preference benchmark, a crowdsourced measure of user taste rather than raw capability, and the distinction matters. By April 25th the full picture was clear enough to quote directly: 1.6 trillion parameters, MIT license, $1.74 per million tokens. On April 28th, DeepSeek cut V4-Pro pricing by a further 75%, temporarily. Someone on LocalLLaMA did the arithmetic on their own stack and reported back, with the specific exhaustion of someone who has just realized their infrastructure decisions need revisiting, that the pricing was "genuinely silly."

If confirmed, here is what this means. A 1.6-trillion-parameter model released under MIT license at $1.74 per million tokens — then discounted to under fifty cents during the promotional window — does not merely compete on price. It restructures the conversation about what frontier AI costs, who can afford to build with it, and what Western lab pricing actually reflects. The 384K output window is not a footnote; it changes what class of task is even addressable via API, opening document-length generation, long-form code synthesis, and multi-step reasoning chains that were previously impractical at scale. The mHC architecture finding matters because it suggests DeepSeek is not simply scaling conventionally — they are doing something structurally different inside the model, which makes the benchmark comparisons harder to interpret and the long-term trajectory harder to predict. The Arena underperformance on user preference is a real signal, not a dismissal: it suggests V4 Pro may optimize for capability over palatability, which is a strategic choice, not an accident. For anyone running inference infrastructure, the LocalLLaMA pricing math is the moment of reckoning — the gap between US and Chinese AI economics has become impossible to attribute purely to regulatory arbitrage or subsidized compute. Something deeper is happening in how they build.

Watch for independent replications of the mHC architecture analysis — if the mechanism holds up under scrutiny, the efficiency story becomes structural rather than incidental. And watch whether the 75% pricing cut becomes permanent; if it does, it stops being a promotion and starts being a market position.

NewsHive monitors these sources continuously. All signal titles above link to the original reporting.

Intelligence by NewsHive. Need help navigating what this means for your business? Contact GeekyBee →