What DeepSeek actually is, and what changed in V4
DeepSeek came out of a Chinese hedge fund and ships MIT-licensed models. What V4 changed, how it measures, and where our copy of it runs.
For most people the name arrived attached to a stock chart. In late January 2025 a Chinese lab almost nobody outside AI research had heard of released a reasoning model called R1, its app went to number one on the US App Store, and Nvidia's share price fell about 18% in a day. Then the cycle moved on and DeepSeek settled into being one of those names you half recognise without being able to say what it is.
Eighteen months later there is a second generation, V4, and it arrived to something close to silence. Since we serve one of its models, here is the plain version: who builds it, what open weights mean for someone who will never run a model themselves, what V4 changed, and where it competes.
A hedge fund's research lab
DeepSeek was founded by Liang Wenfeng in July 2023 and is based in Hangzhou. It is owned and funded by High-Flyer, a Chinese quantitative hedge fund that Liang co-founded in 2015, and it was spun out of High-Flyer's internal AI research group. High-Flyer had been accumulating Nvidia GPUs since 2021 for its own trading work, which is how a hedge fund ended up holding a training cluster.
That origin explains the number that made people sit up. DeepSeek published a training cost of roughly $5.6 million for V3, the base model underneath R1, at a time when its American counterparts were spending hundreds of millions. It also releases its models under the MIT licence, which matters more to a normal user than the price tag ever did.
What open weights get you
When DeepSeek releases a model, it publishes the weights: the billions of numbers that constitute the trained model. Under MIT, anyone may download them, run them, host them, modify them and sell access to them. The industry calls this open weights rather than open source, and the distinction is real. You get the finished model. You do not get the training data or the recipe that produced it.
Two consequences follow. Independent evaluators can host the identical weights and publish comparable measurements, which is why the numbers below exist at all. And the company that trained the model is often not the company answering your question, because the weights run on hardware owned by whoever chose to load them. That second point decides most of the questions people actually have about DeepSeek.
Then V4 arrived
DeepSeek announced V4 on 24 April 2026 as a preview, in two sizes. V4-Pro carries 1.6 trillion total parameters with 49 billion active; V4-Flash carries 284 billion total with 13 billion active. Both are mixture-of-experts models, meaning only a slice of the network fires for any given word, which is how a 284-billion-parameter model can answer at the pace of something far smaller. Both shipped with a 1 million token context as the new default across DeepSeek's services, and both offer Thinking and Non-Thinking modes, so the model can either reason at length before answering or reply directly. The previous generation, deepseek-chat and deepseek-reasoner, was retired on 24 July 2026.
Checkpoints followed on DeepSeek's own news feed through the summer: V4-Flash-0731 on 31 July, the official build replacing the preview; V4-Pro's general release on 13 August; and a separate experimental vision build on 21 August. The two main V4 models are text in, text out.
How it measures, and why those numbers move
Artificial Analysis runs its own evaluations against hosted copies of the weights. Checked on 4 September 2026, it puts V4 Flash 0731 at 52 on its Intelligence Index, 8th of the 112 models in its comparison set, and V4 Pro 0813 at 53, 6th.
Those measurements describe a dated snapshot. A snapshot we took of the same two pages a week earlier had Flash at 107.9 tokens per second and 1.50 seconds to first token; at the later September check the same page said 140.0 tokens per second and 1.19 seconds. Rankings shift as new models enter and evaluations are re-run, so any article quoting a leaderboard position without a date is describing a moment that has already passed, this one included.
With that caveat, the comparison worth making is between the two V4 builds. They sit one index point apart, and Flash produces output at 140 tokens per second against Pro's 60.2. In a chat window that is the difference between an answer that appears and an answer you sit through. The smaller build is the one you would actually want in front of you.
It has a habit worth knowing about. Artificial Analysis flags V4 Flash as very verbose: 210 million tokens generated across its evaluation suite, against a median of 110 million for comparable models. Ask it something simple and you will often get more than you wanted. Where it competes well is sustained reasoning over text you supply, which is also where the verbosity is least annoying.
Why nobody panicked this time
R1 caused alarm. V4 caused a shrug, and an analysis published in May 2026 by the Foreign Affairs Forum sets out why. The lesson that a Chinese lab could reach the frontier had already been absorbed, so there was no discovery to be shocked by. Alibaba's Qwen, ByteDance's Doubao, Moonshot and Z.ai now occupy the same ground, which makes a strong open-weight release ordinary. The V4 announcement disclosed no training cost, where the earlier generation had made a point of it. And the V4 documentation contains no reference to safety measures, red-teaming or risk evaluation, which is a procurement problem in regulated markets regardless of how the model scores.
The same piece adds a structural point that gets less attention than the benchmark tables:
The lab's software innovations were designed to compensate for hardware deficits; as those deficits deepen, the compensation required grows more demanding.
A measured shift in what it will answer
One independent study looks directly at the build we serve. CTGT published LineageEval results on 5 August 2026 comparing the V4-Flash preview against the official 0731 release, using 152 matched pairs of sensitive and control prompts scored by a four-judge panel. Its censorship score on China-sensitive prompts rose from 57.4 to 63.8. On the matched control prompts the same score fell from 25.4 to 19.8. The gap between the two widened rather than everything tightening together, which is what CTGT means by selective.
That is one study from one lab using one method, and CTGT itself is careful to say it cannot be certain of the mechanism and is reporting a measured behavioural change between two builds. It is worth knowing anyway, because 0731 is the exact build sitting in our model picker, and you would rather hear it from us.
Where our copy of it runs
One DeepSeek model is active in our catalog: DeepSeek V4 Flash, the 0731 build, available on the free plan and flagged as recommended. We do not carry V4 Pro and we do not carry the experimental vision build, so this model takes text and text files, with no image or PDF input and no web search. An older V3.2 entry is marked deprecated and already resolves to V4 Flash when selected. DeepSeek publishes a 1 million token context for it; the working window we actually run it at is smaller, and a free account cannot set its reasoning effort at all, because the app picks the level itself and caps its own pick here at medium.
Routing is the question people ask next, and the answer is short. The request never reaches DeepSeek. The entry is marked for restricted routing, which pins it to a fixed allowlist of six hosts (DeepInfra, Fireworks, Google Vertex, Amazon Bedrock, Groq and Azure) and separately blocks model creators' own endpoints along with hosts whose custody chain we cannot audit.
On training, the accurate line is per-model. We do not train on your conversations. Some providers might, depending on which model you choose, and we tell you which ones on the model's information card in the app. At the September 2026 check, six models in our catalogue carried a note there. Most read "Routed via OpenRouter to third-party providers that may briefly retain requests; they do not train on your data", while Grok's says plainly "xAI may use your conversations to train Grok models." DeepSeek V4 Flash is not one of the six.
What it is good for now
DeepSeek has become a normal supplier, which is a stranger outcome than the one people expected in 2025. V4 Flash is quick, wordy and text-only, sits near the top of the open-weight field, and lands within one index point of a sibling more than five times its size while answering more than twice as fast. That combination earns it a place in a picker. It also ships, on that same reading of its documentation, with nothing published about safety testing, and with one independent measurement of selective tightening on China-sensitive questions. Both of those are true at the same time as the first part.
So use it the way its shape suggests: reasoning through a long document, drafting, restructuring an argument, anything where you brought the material and want it worked on quickly. It is a poor choice when the model itself has to be the authority on a contested political question. The thing to carry forward is the licence: a model this capable now ships under terms that let anyone host it, which is why you can use DeepSeek without ever sending anything to DeepSeek.