Hello, AI enthusiasts! Editor Z here from AITech blogZ. We first covered DeepSeek back in February 2025, days after the Chinese startup's R1 model briefly overtook ChatGPT as the most-downloaded free app on the U.S. App Store and wiped roughly $600 billion off Nvidia's market cap in a single trading session. That was a moment, not a conclusion. Since then DeepSeek has shipped three more model generations, gotten banned from government devices in at least 17 U.S. states, and had its headline cost claims picked apart by outside researchers. This is an update on how the technology actually works and where things stand now.
How DeepSeek's Reasoning Models Work
DeepSeek R1, released in January 2025, was built around chain-of-thought reasoning: instead of producing an answer in one pass, the model works through a problem in visible intermediate steps before committing to a final response. That's the same basic idea behind OpenAI's o1, which had shipped a few months earlier. What made R1 notable wasn't the concept but the way DeepSeek got there — its precursor, R1-Zero, was trained to reason using reinforcement learning alone, with no human-labeled step-by-step examples to imitate. The model learned to reason by being rewarded for correct answers and penalized for wrong ones, and it arrived at chain-of-thought behavior on its own. R1 itself added a round of supervised fine-tuning on top of that to fix R1-Zero's readability problems (it would sometimes mix languages mid-answer), producing a more usable model without abandoning the reinforcement-learning core.
The other piece is architectural. DeepSeek's models use a Mixture of Experts (MoE) design: rather than running every parameter for every query, the model routes each token to a small subset of specialized "expert" subnetworks. DeepSeek-V3, the 671-billion-parameter base model R1 was built on, activates only about 37 billion parameters per token. That's why DeepSeek's models could plausibly run and train more cheaply than a dense model of comparable size, even before you get to debates about the actual dollar figures.
The "$5.6 Million" Number, Revisited
The original DeepSeek-V3 paper cited a training cost of roughly $5.576 million for the final pre-training run on a cluster of Nvidia H800 GPUs — a figure widely repeated at the time, including by us, as evidence DeepSeek had matched frontier labs at a fraction of their spend. That number was accurate as far as it went, but it was never the whole picture. It covered only the final training run's GPU-hours at market rental rates; it excluded R&D salaries, the cost of prior experiments and failed runs, data curation, and the capital cost of the underlying hardware cluster itself. Analysts at Bernstein and elsewhere argued the true all-in cost was several times higher, and SemiAnalysis estimated DeepSeek's parent company High-Flyer had access to a compute stack including roughly 10,000 A100s, 10,000 H100s, and 10,000 H800s — hardware worth well over a billion dollars, acquired before the tightest U.S. export controls took effect. The fairer framing: DeepSeek trained efficient models relative to their scale, but "$5.6 million total" was always a narrower claim than the headlines made it sound.
The Model Lineage: V1 Through V4
DeepSeek has kept shipping on a roughly quarterly cadence since R1's debut, and the lineup has moved well past what we covered originally:
| Model | Released | Parameters | What changed |
|---|---|---|---|
| DeepSeek V1 | Jan 2024 | 67B | Dense transformer, first public release |
| DeepSeek V2 | Jun 2024 | 236B | Introduced Multi-Head Latent Attention and MoE routing |
| DeepSeek V3 | Dec 2024 | 671B (37B active) | Reinforcement learning in post-training, GPU efficiency work |
| DeepSeek R1-Zero / R1 | Jan 2025 | 671B (37B active) | Pure-RL reasoning, then RL + supervised fine-tuning |
| DeepSeek V3.2-Exp / V3.2 | Sep–Dec 2025 | 671B (37B active) | DeepSeek Sparse Attention (DSA), cuts long-context compute cost sharply |
| DeepSeek V4, V4-Pro, V4-Flash | Apr 2026 | Undisclosed (MoE) | 1-million-token context window, MIT license, dual "thinking / non-thinking" modes |
| DeepSeek V4-Flash-0731 | Jul 2026 | — | Agent/tool-use upgrade, Responses API and Codex compatibility |
Two things are worth flagging. First, DeepSeek's own R2 — the expected reasoning-focused successor to R1 — still hasn't shipped as of this writing. Reuters reported that CEO Liang Wenfeng held it back because internal benchmarks weren't clearing the bar he wanted, and the company has since folded reasoning capability directly into the V4 line's "thinking mode" rather than releasing a separate R-series model. Second, DeepSeek retired the old deepseek-chat and deepseek-reasoner API aliases after July 24, 2026, so any integration still pointing at those endpoints needs to be repointed at the current V4 model names.
DeepSeek V4 Against Today's Frontier Models
The comparison that mattered in February 2025 was R1 versus o1. The comparison that matters now is DeepSeek's current flagship against GPT-5.x and Gemini 3.1 Pro — and the picture is more mixed than "beats everyone for less."
| Model | License | Coding (Codeforces Elo) | Factual accuracy (SimpleQA-Verified) | API price (input/output per 1M tokens) |
|---|---|---|---|---|
| DeepSeek V4-Pro-Max | MIT (open weights) | 3,206 | 57.9 | ~$0.28 / $0.42 (V3.2 tier pricing carried forward) |
| GPT-5.4 (xHigh reasoning) | Proprietary | 3,168 | higher, unpublished exact figure | Premium tier, no open weights |
| Gemini 3.1 Pro (High) | Proprietary | 91.7 on LiveCodeBench | 75.6 | Premium tier, no open weights |
DeepSeek V4-Pro-Max edges out GPT-5.4 on competitive-programming Elo and posts the highest GPQA Diamond science-reasoning score of any model tested so far (94.3%). But on SimpleQA-Verified — a benchmark specifically designed to catch confident wrong answers — it trails Gemini 3.1 Pro by close to 18 points, meaning it hallucinates more often on short factual questions. The honest summary: DeepSeek's models are genuinely frontier-competitive on reasoning and code, still priced far below the closed-model leaders, and released with fully open MIT-licensed weights on Hugging Face — but "beats OpenAI at everything" was never accurate, and it's less accurate now than it was in February 2025 when the field had fewer data points.
The Part the Original Post Skipped: Bans and Data
What our original coverage left out entirely is what happened after the download numbers made headlines. Within weeks of R1's release, the U.S. Congress's Select Committee on the Chinese Communist Party published findings that the DeepSeek app sent user data to servers in China, contained security vulnerabilities, and applied government-mandated censorship on topics like Tiananmen Square and Taiwan. The Pentagon's Defense Information Systems Agency blocked access on Defense Department networks, the Navy barred its use entirely, and Congress restricted it on staff devices. By 2026, at least 17 states — including Texas (the first, in January 2025), New York, Virginia, and Georgia — had banned DeepSeek from state government devices and networks, and more restrictions have followed at the state level since.
None of that is a verdict on whether the open-weight model itself is safe to self-host and run locally, which sidesteps the data-transmission concern entirely — plenty of developers do exactly that with the Hugging Face weights. It is a real constraint on using DeepSeek's own hosted app or API for anything involving government, enterprise-sensitive, or regulated data, and it's a gap the original post — written the week the story broke, before any of these reviews concluded — had no way to cover.
What It Means Going Forward
The lasting effect of DeepSeek's early-2025 moment wasn't any single model — it was proof that a lab outside the U.S. frontier labs, working with export-restricted hardware, could produce a reasoning model competitive with the best closed systems and then give the weights away under an MIT license. That pressure hasn't let up: DeepSeek has kept releasing state-of-the-art open weights roughly every quarter, and rivals from Alibaba's Qwen line to Meta's Llama have leaned harder into open releases partly in response. On the hardware side, DeepSeek is reportedly developing its own AI chip to reduce dependence on Nvidia, whose China-market H20 chips remain subject to shifting U.S. export rules — a dynamic that continues to unsettle Nvidia's stock on any DeepSeek-related news, echoing the January 2025 selloff.
What hasn't materialized is R2, and what has changed is the skepticism around DeepSeek's own numbers — both are useful correctives to how the story was first reported. The technology is real and the competitive pressure it created is real; the cleanest version of "one small startup outspent Silicon Valley by 96%" was always a simplification, and a year and a half of scrutiny has narrowed it considerably.
-EditorZ
Post a Comment