A three-month preview ends, and two model names end with it

Here's the deal: the V4 that DeepSeek shipped on April 24, 2026 with a "preview" tag flips to general availability today. On the same day, two strings that nearly every developer touching Chinese AI has had hardcoded for the past two years stop working. deepseek-chat and deepseek-reasoner. The official API docs say both are "fully retired and inaccessible" after 15:59 UTC on July 24, 2026. That's 11:59 a.m. ET. If your team has a deploy scheduled for this afternoon, that deploy may be your next incident report.

As preview-to-GA transitions go, this one is unusually blunt. The convention is to keep the old names alive as aliases for a while and quietly redirect them. DeepSeek instead capped the grace period at exactly three months and told everyone it starts throwing 404s after that. During the preview window, deepseek-chat calls were being routed internally to deepseek-v4-flash. Today that bridge gets pulled up. The migration instructions are about as simple as they come: leave the base URL alone, change the model field to deepseek-v4-pro or deepseek-v4-flash. Both the OpenAI ChatCompletions format and the Anthropic API format are still accepted, unchanged.

The numbers explain why anyone cares. V4-Pro is a mixture-of-experts model with 1.6 trillion total parameters and 49 billion active per token. V4-Flash runs 284 billion total, 13 billion active. Both ship with a 1 million token default context and a 384,000 token maximum output. Training consumed more than 32 trillion tokens — more than double the 14.8 trillion that went into V3 at the end of 2024. And all of it sits on Hugging Face under an MIT license: commercial use, modification, redistribution, fine-tuning, no strings. That's about as permissive as open licensing gets.

Then there's the thing DeepSeek has never done in its entire existence: time-of-day pricing. Call the API between 9 a.m. and noon or between 2 p.m. and 6 p.m. Beijing time and you pay exactly double. The company that spent two years being called the industry's price destroyer just invented an expensive hour. Frontier-grade open weights, a million-token window, and surge pricing — all landing on the same Friday.

Where DeepSeek actually stands right now

If you remember the R1 shock of January 2025, the interesting question is why the company went quiet afterward. R1 wiped close to a trillion dollars off Nasdaq market cap in a single session and knocked Nvidia down 17%. It was the moment Liang Wenfeng's lab — spun out of the hedge fund High-Flyer — proved you could get near the frontier for dramatically less money. Then, for eighteen months, DeepSeek shipped only incremental updates: V3.1, V3.2. The gap from V3 to V4 ran 484 days. Over that same stretch, OpenAI cycled through the GPT-5 family and Anthropic went through multiple Claude 4.x generations.

Chinese outlets and analysis shops like ChinaTalk have gradually filled in what happened in that gap. Four things, roughly. First, moving the training framework off Nvidia and onto Huawei Ascend turned out to be far uglier than planned — reporting suggests a fairly serious training failure during chip re-adaptation in mid-2025. Second, there was internal disagreement about training direction, and Liang's requirements were reportedly hard to reconcile at the engineering level. Third, Tencent, ByteDance, and Xiaomi pulled away meaningful talent. Fourth, multimodal generation training got pushed to the back of the queue entirely, squeezed by both compute and cash.

The cash part is not a small detail. DeepSeek opened an external financing window for the first time in mid-April 2026. Until then it ran purely on High-Flyer's own capital, and reporting indicates fundraising dragged because the lab turned down exclusivity terms that large investors wanted. While US frontier labs were closing rounds in the tens of billions, China's best open-source lab was buying GPUs with its parent company's trading profits.

Even so, V4 carries the same DeepSeek fingerprint: solve it with engineering. The model card describes a hybrid attention stack — compressed sparse attention (CSA) mixed with high-compression attention (HCA) — that cuts inference FLOPs per token to 27% of V3.2's at the 1M-token point. Reading the same long document costs roughly a quarter of the compute. On top of that, manifold-constrained hyper-connections strengthen cross-layer signal flow, and training used the Muon optimizer. Inference artifacts ship in mixed precision: expert parameters in FP4, everything else in FP8. Huawei announced day-zero full support for V4 on Ascend supernodes and said post-training for V4-Pro completed on an Ascend 910C cluster of at least 1,000 cards. But per ChinaTalk, pretraining itself still happened on Nvidia hardware, and the domestic-chip adaptation is aimed mainly at inference. The "train and serve entirely without Nvidia" story is not finished yet.

1.6 trillion, 284 billion, and a price list that checks the clock

V4 splits into two models. Pro where you need capability, Flash where you need volume. Both expose three reasoning modes — Non-Think, Think High, Think Max — so you dial the thinking budget inside a single model instead of swapping endpoints. The model card claims Think Max "substantially raises the knowledge ceiling of open-source models," which only means something once you look at the benchmark numbers. V4-Pro posts MMLU 90.1% (5-shot), GSM8K 92.6% (8-shot), HumanEval 76.8% (0-shot). V4-Flash in thinking mode posts MMLU-Pro 86.4, LiveCodeBench 91.6, and 78.7 on MRCR, the long-context retrieval benchmark measured at the 1M-token point. Community reporting has V4-Pro at 80.6% on SWE-bench Verified — 0.2 points behind Claude Opus 4.6 at 80.8% — but that figure has not been cross-confirmed in DeepSeek's own materials, so treat it as unverified.

The pricing is where this announcement gets spicy. In May, DeepSeek attached discounts of up to 75% to the V4 API, triggering a price war that dragged ByteDance and Tencent down with it. Two months later the company reversed and decided to charge double when demand concentrates. The stated reason is "better resource allocation and stronger service stability." Translated: it will hold the off-peak floor and use price to push peak-hour load somewhere else. It's time-of-use electricity billing, ported to tokens.

Item V4-Flash (off-peak) V4-Flash (peak) V4-Pro (off-peak) V4-Pro (peak)
Input · cache hit (1M tokens) $0.0028 2x $0.003625 2x
Input · cache miss (1M tokens) $0.14 2x $0.435 $0.87
Output (1M tokens) $0.28 $0.56 $0.87 $1.74
Total / active params 284B / 13B — 1.6T / 49B —
Context / max output 1M / 384K tokens — 1M / 384K tokens —
License MIT (as of preview) — MIT (as of preview) —

In yuan terms, V4-Pro output goes from 6 yuan per million tokens to 12 at peak, and cache-miss input goes from 3 to 6. Peak hours are 09:00–12:00 and 14:00–18:00 Beijing time, seven hours a day total. For US teams that maps to roughly 9 p.m.–midnight and 2 a.m.–6 a.m. ET — meaning American business hours land squarely in the cheap window, while European afternoons and most of Asia's workday do not. Nightly batch jobs in China get hit; nightly batch jobs in California don't.

One caveat worth flagging. The April preview weights for both V4-Pro and V4-Flash went up on Hugging Face and ModelScope under MIT, runnable on vLLM, SGLang, and Transformers. Whether today's GA transition refreshes those weights with a new checkpoint, and whether the license terms carry over unchanged, is not confirmed as of writing. There's also a community claim that V4-Flash gains image input, which is not in any official document. If you're planning to self-host, check the commit hash on today's model card yourself.

Who's actually winning here

DeepSeek gets two things. One is control over serving cost. Running a 1.6T MoE at a million tokens of context is brutal on KV cache memory no matter how far CSA and HCA push FLOPs down. Double the peak price and either revenue goes up or load goes down — both outcomes work for DeepSeek. The other is a cleaner product line. Names built around "does it reason or not" stopped making sense the moment three thinking modes moved inside one model. Splitting by size instead — Pro and Flash — gives them a naming scheme they can reuse for V5 and V6.

Huawei quietly takes the biggest prize. A reference deployment where a 1.6T model completed post-training on an Ascend 910C cluster does enormous credibility work for Ascend as an Nvidia substitute inside China. And the tooling DeepSeek built along the way is all open source: the MXFP4 low-precision format, TileLang for multi-hardware compilation, the MegaMoE fused kernels that actually run on Ascend. The next Chinese lab that wants to migrate has a paved road.

Chinese cloud and API resellers are in an awkward spot. Off-peak prices hold, so the floor of the discount war doesn't move — but if your input cost doubles for seven hours a day and you sell flat-rate plans, your margin gets squeezed. As SCMP noted, some read this as a signal that the cost-leadership strategy that defined the last two years has hit a wall against real infrastructure demand. For resellers, shoving traffic into off-peak batch processing is basically the only defense available.

Western frontier labs can't ignore the sticker. V4-Flash output runs $0.28 per million tokens, and US commercial models claiming comparable capability generally sit several multiples above that. On top of that, V4's weights are MIT, so "if you charge too much I'll just run it myself" is a live option rather than a bluff. The real effect isn't the price list itself — it's the anchor it sets in every enterprise negotiation.

Enterprises with data-residency constraints get an extra calculation. Plenty of organizations — in Korea, the EU, regulated US sectors — can't send data to a Chinese API at all. MIT open weights change that conversation completely, because you can run the thing inside your own perimeter. The practical catch is hardware: 1.6T Pro is heavy to serve even at FP4 mixed precision, so the realistic on-prem candidate is Flash at 284B total and 13B active.

Preview-to-GA bridges: the ones that held and the ones that collapsed

Killing old model names when the preview label comes off is a ritual the industry has repeated for years. On the success side, OpenAI phased out the text-davinci-003 completion family across 2023–2024 and consolidated everything onto the chat completions format. There were migration guides, there was a grace period, and the ecosystem came out ahead because the API surface got simpler. DeepSeek accepting both the OpenAI and Anthropic request formats while telling you to change only the model string is a direct descendant of that lesson.

The failure cases are just as instructive. When Google moved the PaLM API to Gemini across 2023–2024 and changed endpoints and SDKs at the same time, and when various clouds promoted preview APIs to GA while quietly adjusting response schemas, a lot of production quietly broke. The common thread: the announcement said "only the name changes," but the tokenizer or the default parameters changed too. Since nobody has confirmed that DeepSeek's preview checkpoint and GA checkpoint are identical, any team with heavily tuned prompts should run an A/B regression before trusting it.

DeepSeek's own history is the most relevant precedent. Right after R1 launched in 2025, the API got hammered hard enough that new signups were suspended for days. The fix DeepSeek chose then was not a price increase — it was a steep off-peak discount, 50–75% off between 16:30 and 00:30 UTC. Carrot, not stick. Eighteen months later, facing the same congestion problem, it reached for the stick. A company using the exact opposite tool on the exact same problem is also an admission: the discounts alone never moved enough daytime load.

How the competition responds

The most immediate pressure comes from Moonshot AI. Kimi K3 hit the API on July 16 with 2.8 trillion total parameters, a 1M-token context, and native multimodal input — and Moonshot has already announced full open weights for July 27, three days after DeepSeek's GA. On raw parameter count K3 is 1.75x V4-Pro, and multimodal is a lane DeepSeek explicitly skipped this cycle. Observers including Nathan Lambert have called K3 the strongest open model ever released, ahead of everything except the top Claude and GPT tiers. The era where DeepSeek automatically was the face of Chinese open source is over.

The timing is awkward in another way. WAIC 2026 ran in Shanghai from July 17 to 20 — over 1,000 exhibitors and 300-plus global product debuts — and DeepSeek's GA landed after the show closed. Moonshot, meanwhile, dropped K3 the day before the doors opened. The headline at the biggest annual stage in Chinese AI went to Moonshot, and DeepSeek is handling its stable release and model-name retirement in a quiet week.

ByteDance (Doubao) and Alibaba (Qwen) will likely counter from a different angle. Both own their own clouds, so neither has the same urgency to push serving cost onto customers through a rate card. The easiest play is marketing: "no time-of-day surcharge here." Given that both followed DeepSeek down when it opened the May price war with 75% off, this time they can collect the benefit simply by not following it up. Qwen's real strength is the breadth of its open-weights lineup — small models through large — so expect it to defend on the fine-tuning ecosystem rather than go head-to-head with a 1.6T MIT model.

The Western response arrives slower. OpenAI and Anthropic already pushed effective unit costs way down with batch APIs and prompt caching, and DeepSeek's cache-hit pricing ($0.0028–$0.003625 per million tokens, effectively free) is an extension of the same caching arms race. Neither is likely to cut list prices immediately, but "V4-Flash is twenty-eight cents" is going to show up in enterprise negotiations constantly. On the open-weights front, the open question is what Meta and Mistral do. In a market where 2.8T (K3) and 1.6T (V4-Pro) ship under MIT-grade licenses, a mid-size model with a restrictive license has very little room to stand.

So what actually changes

If you're a developer, today has exactly one mandatory task: grep your entire codebase for deepseek-chat and deepseek-reasoner and replace them with deepseek-v4-flash or deepseek-v4-pro. That includes environment variables, CI secrets, LangChain and LlamaIndex config files, and values hardcoded inside notebooks. Then look at scheduling. Latency-tolerant work — batch embeddings, bulk summarization — costs half as much if it runs outside Beijing business hours, which for US teams is most of the workday anyway. If you run a real-time chatbot serving Asian users, your blended unit cost just went up, and the budget should reflect it.

If you're an investor, the signal here is direction, not price. For two years the dominant narrative was that Chinese AI wins on cost — and the company that authored that narrative just announced, for the first time, that it charges more when demand concentrates. That means either inference demand has outrun supply, or compute procurement is still the binding constraint, or both. Pair it with the April decision to open external financing and the second reading gets heavier. Any model of Chinese AI valuations built purely on token unit price needs revisiting.

If you're a regular user, honestly not much changes. The DeepSeek app and web chat stay free, and the interface doesn't advertise which model is running underneath. What you may notice is that a 1M-token context is now the default across all official services — throw several long PDFs in at once and keep the conversation going without the model forgetting the beginning. MRCR 78.7 is the number that measures this, and it doesn't mean perfect: read it as finding the right thing about four times out of five, even out at a million tokens.

If you build foundation models, the open-weights part is the pressure. A model trained on 32 trillion tokens is now sitting on Hugging Face under MIT, free, permanently. Selling a comparable in-house model at a price requires a new argument — data residency, domain tuning, support guarantees, language-specific quality — because "ours is better" is no longer sufficient on its own. On the flip side, teams with sovereignty requirements just got a genuinely deployable base: at 13B active parameters, V4-Flash is servable on hardware people actually own.

🥄 Three Things You're Probably Wondering

— So what does this mean for me? Not much directly, unless you call the DeepSeek API. But a lot of AI products quietly blend DeepSeek models in on the backend, and if one of your vendors missed the model-name swap, you might see weird latency or dead requests starting this afternoon.

— Why now, the week after WAIC? DeepSeek is famously bad at marketing timing. One reading is that this is just a few days' slip from the "mid-July" it signaled on June 30. Another is that it deliberately avoided a head-on collision with Kimi K3. The company hasn't said, so pick your theory carefully.

— Is this ahead of Kimi K3 or behind it? On parameter count (2.8T vs 1.6T) and multimodal support, K3 is ahead. V4 counters with an efficiency design that cuts inference FLOPs to 27%, three months of proven API stability, and much lower serving cost. K3's weights don't land until July 27, so any real ranking has to wait until next week.

Sources

Numbers are as of announcement and may change.