Introduction
After more than three months in preview, the flagship model of DeepSeek (杭州深度求索人工智能基础技术研究有限公司) has reached general availability (see Appendix A, Glossary, for technical terms introduced in this essay, and Appendix B for the full translation and meaning of DeepSeek’s name in Chinese). The Hangzhou-based company formally released DeepSeek-V4-Pro-0813 across its app, web interface, and API on August 13, 2026, closing out a preview window that began on April 24, 2026, and following the smaller V4-Flash model into production (DeepSeek, 2026b; Reuters, 2026).
The release is notable for three reasons: the general-availability build posts large gains over its April preview on agentic and coding benchmarks; DeepSeek is simultaneously retiring its heavily discounted promotional pricing in favor of a peak/off-peak structure that takes effect August 16, 2026; and even after that increase, V4 Pro remains far cheaper than the leading Western frontier models it is benchmarked against.
This article reviews the transition from preview to general availability, compares the new model’s performance and price against both its preview predecessor and competing frontier systems, and examines the organizational and technical strengths — and the evidence-based research lineage behind them — that allow a comparatively small Chinese lab to field models of this caliber.
From Preview to General Availability
DeepSeek staged the V4 rollout deliberately. The V4 family first appeared on April 24, 2026 as open-weight previews, with both the Pro and Flash variants released under the permissive MIT license (DeepSeek, 2026a). DeepSeek-V4-Pro is a mixture-of-experts (MoE) model with 1.6 trillion total parameters and roughly 49 billion active per token, while V4-Flash is a 284-billion-parameter model with 13 billion active parameters; both support a one-million-token context window — the amount of text the model can hold in view at once, effectively its working memory — and were pre-trained on more than 32 trillion tokens (DeepSeek-AI, 2026).
These figures are easiest to grasp through what a mixture-of-experts (MoE) model actually is. Imagine a hospital staffed by hundreds of specialists and a single triage nurse: the nurse reads each incoming case and dispatches it to only the few specialists whose expertise matches, so the hospital holds far more combined knowledge than any one patient requires. An MoE model works the same way: rather than one network in which every parameter fires on every input, it divides its capacity across many specialized sub-networks (the “experts”) and uses a small router — the triage nurse — to send each token (a small chunk of text, roughly a word or part of one) to only a few of them. The result is sparse computation: of V4-Pro’s 1.6 trillion parameters, only about 49 billion — roughly 3% — are active for any single token, while different tokens activate different experts so the full capacity is still used across a conversation (DeepSeek-AI, 2026). V4-Flash is sparser still, activating about 13 billion of its 284 billion parameters, roughly 4.6% (DeepSeek-AI, 2026).
This decoupling of a model’s size from its running cost is the core of DeepSeek’s efficiency advantage, though it should be understood precisely: sparsity cuts the compute each token requires, not the memory needed to host the full weights, which for V4-Pro amount to an 865 GB download (Farnschläder, 2026). The flagship uses a hybrid attention design — combining compressed sparse attention and heavily compressed attention — that DeepSeek reports cuts inference compute at one-million-token context to 27% of what V3.2 required and reduces the key-value cache to just 10%, an efficiency critical to serving a million-token window economically (DeepSeek-AI, 2026; Farnschläder, 2026).
The sequencing drew unusual attention. V4-Flash, the cheaper model, graduated to official status first, on July 31, 2026, and in independent tests it sometimes outperformed the April preview of V4-Pro — an inversion, given that Pro is positioned as the company’s more capable product (Reuters, 2026; Tolomia, 2026a). That outcome suggested DeepSeek had improved its technology rapidly between the preview and the official launch, and the Pro release was “closely watched” as a result (Reuters, 2026). The general-availability build, DeepSeek-V4-Pro-0813, is oriented toward agentic workloads — tasks in which a model uses tools, executes code, and completes multi-step work — and adds three selectable reasoning-effort levels (low, high, and max) alongside native support for the OpenAI Responses API and a one-click Codex integration (DeepSeek, 2026b; Reuters, 2026).
Performance: General Availability vs. Preview
DeepSeek’s own benchmark release for the 0813 build shows the largest improvements precisely where the April preview was weakest. The preview scored 12.8 on the agentic coding benchmark DeepSWE; the general-availability build scores 62.7, a gain of roughly 50 points (DeepSeek, 2026b). Comparable jumps appear across the agent-focused suite: CyberGym rises from 52.7 to 83.3, DSBench-Hard from 31.1 to 67.2, DSBench-FullStack from 41.8 to 71.1, and Terminal Bench 2.1 from 72.1 to 87.9 (DeepSeek, 2026b). Table 1 presents the full comparison.
Table 1
DeepSeek V4-Pro Agentic and Coding Benchmarks: General Availability Versus Preview
| Benchmark | V4-Pro GA (0813) | V4-Pro Preview | Change |
|---|---|---|---|
| Terminal Bench 2.1 | 87.9 | 72.1 | +15.8 |
| DeepSWE | 62.7 | 12.8 | +49.9 |
| NL2Repo | 61.5 | 38.5 | +23.0 |
| CyberGym | 83.3 | 52.7 | +30.6 |
| Toolathlon-Verified | 74.1 | 55.9 | +18.2 |
| AutomationBench (Public) | 31.8 | 12.8 | +19.0 |
| DSBench-FullStack | 71.1 | 41.8 | +29.3 |
| DSBench-Hard | 67.2 | 31.1 | +36.1 |
| Humanity's Last Exam (w/o / w tools) | 42.7 / 60.0 | 37.7 / 48.2 | +5.0 / +11.8 |
| Agents' Last Exam | 25.7 | 16.5 | +9.2 |
Note. Scores are as reported in DeepSeek’s official release announcement (DeepSeek, 2026b) and are vendor-reported rather than independently confirmed. GA = general availability; w/o = without tools; w = with tools.
These figures illustrate a genuine maturation rather than a marginal polish: the preview was effectively unfinished on agentic tasks, and the gap between “preview” and “product” is the story of the release (DeepSeek, 2026b). It is worth keeping in mind that all of these numbers originate from DeepSeek itself and have not yet been independently reproduced for the 0813 build specifically.
Performance: V4 Pro Against the Frontier
Against contemporary frontier systems, the general-availability V4 Pro is competitive in agentic and coding domains while trailing the leading proprietary models on the hardest knowledge tests, a profile consistent with DeepSeek’s stated position as an open-weights lab that trails the absolute frontier by a matter of months (Farnschläder, 2026).
On Terminal Bench 2.1, V4-Pro-0813’s 87.9 places it effectively at the top of the models DeepSeek compared, ahead of Anthropic’s Claude Opus 4.8 (85.0) and within a fraction of a point of Kimi K3 (88.3) and Anthropic’s Fable 5 (88.0) (DeepSeek, 2026b). On CyberGym, its 83.3 edges Fable 5 (83.1), Kimi K3 (80.0), and Opus 4.8 (78.3). On the software-engineering benchmark DeepSWE, however, V4 Pro’s 62.7 trails Fable 5 (70.0) and Kimi K3 (67.5), though it leads Opus 4.8 (58.0). The pattern reverses on the most knowledge-intensive evaluation, Humanity’s Last Exam, where V4 Pro’s 42.7 (without tools) sits behind Fable 5 (53.3) and Opus 4.8 (49.8) (DeepSeek, 2026b). Compared with OpenAI’s GPT-5.5, third-party comparisons put V4 Pro behind on terminal-heavy coding (Terminal-Bench 2.0: 67.9 vs. 82.7) but ahead on long-context retrieval (MRCR 1M: 83.5), while trailing on graduate-level reasoning (GPQA Diamond: 90.1 vs. 93.6) (Farnschläder, 2026).
The takeaway is that V4 Pro now occupies the top tier of open models and is genuinely competitive with proprietary systems on agentic coding, while remaining a step behind the leading closed models on the hardest reasoning and knowledge evaluations (DeepSeek, 2026b; Farnschläder, 2026).
Price: From Promotion to Peak/Off-Peak
DeepSeek’s pricing history for V4 Pro runs against the grain of the “cheapest at all costs” reputation it earned with R1 in early 2025 (Woods, 2025). The V4 preview launched with a 75% promotional discount, bringing V4-Pro to $0.435 per million input tokens (cache miss) and $0.87 per million output tokens — rates that were still in effect when the general-availability build shipped (DeepSeek, 2026c; Tolomia, 2026b). With general availability, the company replaced that flat promotional rate with a peak/off-peak structure, effective 16:00 UTC on August 16, 2026 (DeepSeek, 2026c). Peak hours are 01:00–04:00 and 06:00–10:00 UTC; off-peak rates are half the peak rates (DeepSeek, 2026c). The design borrows the logic of time-of-use electricity billing: charge more when demand is highest, in order to shift usage toward quieter hours. Table 2 summarizes the resulting rates and compares them with leading competitors.
Table 2
DeepSeek V4 and Competitor API Pricing per Million Tokens
| Model and period | Input, cache miss ($/1M) | Output ($/1M) |
|---|---|---|
| V4-Pro (peak) | 1.32 | 3.96 |
| V4-Pro (off-peak) | 0.66 | 1.98 |
| V4-Pro (preview flat rate) | 0.435 | 0.87 |
| V4-Flash (peak) | 0.44 | 1.32 |
| V4-Flash (off-peak) | 0.22 | 0.66 |
| GPT-5.5 | 5.00 | 30.00 |
| GPT-5.5 Pro | 30.00 | 180.00 |
| Claude Opus 4.7 | 5.00 | 25.00 |
| Google Gemini 3.1 Pro | 2.00 | 12.00 |
| Anthropic Fable 5 | — | 50.00 |
Note. DeepSeek pricing is from its official documentation (DeepSeek, 2026c); competitor pricing is as reported by Werth (2026), Farnschläder (2026), and Tolomia (2026b). $/1M = U.S. dollars per one million tokens; cache miss = newly submitted input text that the model has not previously cached.
The increase is steep in percentage terms — Quartz calculated rises ranging from 50% to more than 1,100% depending on model, token type, and time of day, with V4-Pro output tokens moving from the $0.87 flat rate to $3.96 at peak (Tolomia, 2026b). DeepSeek framed the change as a way “to allocate resources more reasonably,” with the tiered structure intended to shift developer workloads toward less congested periods (Tolomia, 2026b). Yet even at peak rates, V4-Pro’s $3.96 output price remains well below GPT-5.5 ($30.00), the prior-generation Claude Opus 4.7 ($25.00), and Anthropic’s Fable 5 ($50.00) (Tolomia, 2026b; Werth, 2026). Off-peak, V4-Pro undercuts Google’s Gemini 3.1 Pro ($12.00 output) as well (Werth, 2026). The discount has narrowed, but DeepSeek retains a substantial cost advantage.
What Makes DeepSeek Capable: Strengths as a Model Developer
DeepSeek’s ability to field frontier-competitive models at these prices is not a single bet but a stack of reinforcing advantages: patient capital, an efficiency-first research program, open-weight distribution, and an unusually flat talent model.
Patient capital and an unusual origin. DeepSeek was founded in July 2023 by Liang Wenfeng as a research arm of High-Flyer, a Chinese quantitative hedge fund that had already used AI for trading and had begun accumulating Nvidia GPUs as early as 2021 (Mok, 2025). Because High-Flyer funds the lab itself, DeepSeek operates without the near-term revenue pressure that venture-backed rivals face, which both underwrites its long-horizon goal of artificial general intelligence and lets it treat open-weight releases as a distribution strategy rather than a sacrifice (Mok, 2025; Woods, 2025).
Efficiency as the core competency. DeepSeek’s signature contribution is getting frontier-adjacent results with far fewer resources, a discipline forced in part by U.S. export controls on advanced chips (Mok, 2025; Woods, 2025). Its technical reports document a series of efficiency innovations: multi-head latent attention, which compresses the key-value cache; a mixture-of-experts design that activates only a small fraction of parameters per token; FP8 mixed-precision training; and, in R1, a reinforcement-learning objective (Group Relative Policy Optimization) plus distillation of larger models into smaller ones (DeepSeek-AI, 2024, 2025; Woods, 2025). The V4 technical report extends this lineage with a hybrid attention scheme aimed at making a million-token context tractable to serve (DeepSeek-AI, 2026). Liu (2025) summarizes the pattern: DeepSeek’s breakthroughs are in “achieving greater efficiency — getting good results with fewer resources,” largely through sparsity and memory compression.
Cost discipline, with a caveat. The DeepSeek-V3 technical report disclosed a $5.576 million final training cost, achieved through “optimized co-design of algorithms, frameworks, and hardware” across 2.788 million GPU-hours on 2,048 Nvidia H800 GPUs (DeepSeek-AI, 2024; Mok, 2025). That headline number covers only the final successful training run and explicitly excludes prior research and ablation experiments; independent analysts have estimated DeepSeek’s cumulative hardware investment at well over $500 million (Mok, 2025). The efficiency is real, but the “$6 million model” framing understates the total research investment behind it (Mok, 2025).
Open weights as strategy. Since R1, DeepSeek has shipped most models under the MIT license, which allows free commercial use, modification, and redistribution, and which lets the global research community independently verify the models rather than take the vendor’s word (DeepSeek, 2026a; Woods, 2025). The V4 family continues this practice, with open weights distributed through Hugging Face (DeepSeek, 2026a).
A flat, bottom-up talent model. DeepSeek hired largely new graduates and doctoral students from top Chinese universities into a deliberately flat organization in which projects were distributed “naturally” and shaped by what researchers themselves wanted to pursue — an organizational posture observers likened to Silicon Valley at its most effective (Mok, 2025). This structure, combined with founder Liang Wenfeng’s stated conviction that China’s gap was one of confidence and organization rather than capital, is widely cited as central to the lab’s pace of innovation (Mok, 2025).
The Evidence Base Beneath DeepSeek’s Methods
The techniques that make DeepSeek’s models efficient are not invented in isolation; they extend a documented body of peer-reviewed and foundational research. Tracing that lineage makes explicit whose work DeepSeek’s scientists are applying, and situates the company’s “experience-based” engineering inside an evidence-based tradition.
Mixture-of-experts. The sparse-activation design at the heart of V4-Pro — 1.6 trillion parameters of which only about 49 billion — roughly 3% — are active per token (DeepSeek-AI, 2026) — descends directly from Jacobs, Jordan, Nowlan, and Hinton’s (1991) “adaptive mixtures of local experts,” the original peer-reviewed formulation of routing each input to a subset of specialized subnetworks, and from Shazeer et al.’s (2017) sparsely-gated mixture-of-experts layer, which demonstrated that the approach could scale to thousands of experts.
Attention and its compression. Attention is how a transformer decides which parts of its input to focus on — a spotlight over the text. The transformer architecture everything else builds on was introduced by Vaswani et al. (2017) in “Attention Is All You Need.” DeepSeek’s multi-head latent attention — which compresses the key-value cache that long-context inference must store — extends the memory-reduction line begun by Shazeer’s (2019) multi-query attention and Ainslie et al.’s (2023) grouped-query attention. The efficient attention kernels underlying DeepSeek’s FlashMLA build on Dao et al.’s (2022) FlashAttention.
Low-precision training. DeepSeek-V3’s FP8 training — storing and computing numbers in 8 bits instead of the usual 16 or 32, which shrinks memory use and speeds up training at a small cost in precision — draws on Micikevicius et al.’s (2022) work on FP8 formats for deep learning.
Reasoning via reinforcement learning. R1’s Group Relative Policy Optimization — the objective behind its reasoning ability — extends Schulman et al.’s (2017) proximal policy optimization, while the step-by-step reasoning it elicits was formalized by Wei et al.’s (2022) chain-of-thought prompting.
Distillation and efficiency. The practice of compressing large models into smaller, cheaper ones — a large “teacher” model transferring its skill to a smaller “student” — traces to Hinton, Vinyals, and Dean’s (2015) knowledge distillation. The broader question of how much compute a model actually needs is framed by Kaplan et al.’s (2020) scaling laws and by Hoffmann et al.’s (2022) compute-optimal training analysis, which showed that many large models had been trained with more parameters than their compute budget justified — the exact inefficiency DeepSeek’s design philosophy targets.
This lineage matters because it shows DeepSeek’s edge is less a single breakthrough than an unusually disciplined application of a shared, evidence-based research corpus: the same foundation Western labs build on, applied under resource constraints that reward efficiency (Liu, 2025; Mok, 2025). The company’s technical reports do not merely report results; they are research oriented documents that name and extend this prior work (DeepSeek-AI, 2024, 2025, 2026).
Conclusion
DeepSeek-V4-Pro-0813 closes a four-month preview by turning a genuinely unfinished agentic model into a competitive one: near-top scores on terminal and security-automation benchmarks, a still-wide cost advantage even at its new peak rates, and a million-token context window made economically serviceable through architectural efficiency (DeepSeek, 2026b, 2026c; DeepSeek-AI, 2026). The simultaneous move off promotional pricing signals a company transitioning from a market-shaking price war toward a sustainable commercial posture — one backed by a $7.4 billion first outside funding round in June 2026 and reported plans for a follow-on round at roughly a $74 billion valuation (Reuters, 2026; Tolomia, 2026a). The underlying reasons DeepSeek can keep shipping such models — patient capital, efficiency-first research, open weights, a flat talent organization, and a command of the foundational research literature — are structural, not incidental, and explain why a comparatively small lab continues to set the terms of the open-model frontier.
References
Ainslie, J., Lee-Thorp, J., de Jong, M., Zemlyanskiy, Y., Lebrón, F., & Sanghai, S. (2023). GQA: Training generalized multi-query transformer models from multi-head checkpoints. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing (EMNLP 2023). https://arxiv.org/abs/2305.13245
Dao, T., Fu, D. Y., Ermon, S., Rudra, A., & Ré, C. (2022). FlashAttention: Fast and memory-efficient exact attention with IO-awareness. In Advances in Neural Information Processing Systems 35 (NeurIPS 2022). https://arxiv.org/abs/2205.14135
DeepSeek. (2026a, April 24). DeepSeek V4 preview release [News release]. DeepSeek API Docs. https://api-docs.deepseek.com/news/news260424
DeepSeek. (2026b, August 13). DeepSeek-V4-Pro GA release [News release]. DeepSeek API Docs. https://api-docs.deepseek.com/news/news260813
DeepSeek. (2026c). Models & pricing. DeepSeek API Docs. Retrieved August 17, 2026, from https://api-docs.deepseek.com/quick_start/pricing
DeepSeek-AI. (2024). DeepSeek-V3 technical report (arXiv:2412.19437). arXiv. https://arxiv.org/abs/2412.19437
DeepSeek-AI. (2025). DeepSeek-R1: Incentivizing reasoning capability in LLMs via reinforcement learning (arXiv:2501.12948). arXiv. https://arxiv.org/abs/2501.12948
DeepSeek-AI. (2026). DeepSeek-V4: Towards highly efficient million-token context intelligence (arXiv:2606.19348). arXiv. https://arxiv.org/abs/2606.19348
Farnschläder, T. (2026). GPT-5.5 vs DeepSeek V4: Which frontier model is right for you? DataCamp. https://www.datacamp.com/blog/deepseek-v4-vs-gpt-5-5
Hinton, G., Vinyals, O., & Dean, J. (2015). Distilling the knowledge in a neural network (arXiv:1503.02531). arXiv. https://arxiv.org/abs/1503.02531
Hoffmann, J., Borgeaud, S., Mensch, A., Buchatskaya, E., Cai, T., Rutherford, E., de Las Casas, D., Hendricks, L. A., Welbl, J., Clark, A., Hennigan, T., Noland, E., Millican, K., van den Driessche, G., Damoc, B., Guy, A., Osindero, S., Simonyan, K., Elsen, E., … Sifre, L. (2022). Training compute-optimal large language models (arXiv:2203.15556). arXiv. https://arxiv.org/abs/2203.15556
Jacobs, R. A., Jordan, M. I., Nowlan, S. J., & Hinton, G. E. (1991). Adaptive mixtures of local experts. Neural Computation, 3(1), 79–87. https://doi.org/10.1162/neco.1991.3.1.79
Kaplan, J., McCandlish, S., Henighan, T., Brown, T. B., Chess, B., Child, R., Gray, S., Radford, A., Wu, J., & Amodei, D. (2020). Scaling laws for neural language models (arXiv:2001.08361). arXiv. https://arxiv.org/abs/2001.08361
Liu, T. (2025, January 29). DeepSeek: How a small Chinese AI company is shaking up US tech heavyweights. The Conversation. https://theconversation.com/deepseek-how-a-small-chinese-ai-company-is-shaking-up-us-tech-heavyweights-248434
Micikevicius, P., Stosic, D., Burgess, N., Cornea, M., Dubey, P., Grisenthwaite, R., Ha, S., Heinecke, A., Judd, P., Kamalu, J., Mellempudi, N., Oberman, S., Shoeybi, M., Siu, M., & Wu, H. (2022). FP8 formats for deep learning (arXiv:2209.05433). arXiv. https://arxiv.org/abs/2209.05433
Mok, C. (2025). Taking stock of the DeepSeek shock. Stanford Cyber Policy Center, Freeman Spogli Institute for International Studies. https://cyber.fsi.stanford.edu/publication/taking-stock-deepseek-shock
Reuters. (2026, August 13). DeepSeek releases official V4 Pro model as it steps up expansion. https://www.reuters.com/world/china/deepseek-releases-official-v4-pro-model-it-steps-up-expansion-2026-08-13/
Schulman, J., Wolski, F., Dhariwal, P., Radford, A., & Klimov, O. (2017). Proximal policy optimization algorithms (arXiv:1707.06347). arXiv. https://arxiv.org/abs/1707.06347
Shazeer, N. (2019). Fast transformer decoding: One write-head is all you need (arXiv:1911.02150). arXiv. https://arxiv.org/abs/1911.02150
Shazeer, N., Mirhoseini, A., Maziarz, K., Davis, A., Le, Q., Hinton, G., & Dean, J. (2017). Outrageously large neural networks: The sparsely-gated mixture-of-experts layer. In Proceedings of the 5th International Conference on Learning Representations (ICLR 2017). https://arxiv.org/abs/1701.06538
Tolomia, C. (2026a, August 13). DeepSeek is officially launching its flagship AI model after months in preview. Quartz. https://qz.com/deepseek-v4-pro-official-launch-081326
Tolomia, C. (2026b, August 13). DeepSeek is raising AI developer access prices by up to 1,100% starting Sunday. Quartz. https://qz.com/deepseek-api-price-increase-v4-peak-off-peak-081326
Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., Kaiser, Ł., & Polosukhin, I. (2017). Attention is all you need. In Advances in Neural Information Processing Systems 30 (NeurIPS 2017). https://arxiv.org/abs/1706.03762
Wei, J., Wang, X., Schuurmans, D., Bosma, M., Ichter, B., Xia, F., Chi, E., Le, Q., & Zhou, D. (2022). Chain-of-thought prompting elicits reasoning in large language models. In Advances in Neural Information Processing Systems 35 (NeurIPS 2022). https://arxiv.org/abs/2201.11903
Werth, T. B. (2026). DeepSeek V4 is here: How it compares to ChatGPT, Claude, Gemini. Mashable. https://mashable.com/article/deepseek-v4-preview-comparison-chatgpt-claude-gemini
Woods, A. (2025, January 28). DeepSeek: What you need to know. MIT CSAIL Alliances. https://cap.csail.mit.edu/research/deepseek-what-you-need-know
Note on Sources
The release dates, pricing tables, and benchmark figures in this article were transcribed directly from DeepSeek’s official announcements and API documentation, and cross-checked against Reuters and Quartz reporting. The three DeepSeek technical reports are arXiv preprints; their DOIs (10.48550/arXiv.2412.19437, 10.48550/arXiv.2501.12948, and 10.48550/arXiv.2606.19348) were verified to resolve against DataCite — the registry that mints arXiv DOIs, rather than Crossref. The foundational-research references fall into two categories, distinguished deliberately: (a) peer-reviewed works — Jacobs et al. (1991) in Neural Computation (DOI 10.1162/neco.1991.3.1.79 verified against Crossref) and the NeurIPS/ICLR/EMNLP conference papers by Vaswani et al. (2017), Shazeer et al. (2017), Dao et al. (2022), Wei et al. (2022), and Ainslie et al. (2023); and (b) canonical preprints that have shaped practice without formal journal review — Schulman et al. (2017), Hinton et al. (2015), Kaplan et al. (2020), Hoffmann et al. (2022), Micikevicius et al. (2022), and Shazeer (2019). All arXiv identifiers and author lists were verified against DataCite/arXiv records; machine-learning conference papers (NeurIPS, ICLR, EMNLP) generally lack Crossref DOIs and are cited via their arXiv identifiers. Because this article covers a product released days ago, the release-specific facts rest on primary vendor and news sources rather than peer-reviewed literature; the technical and organizational claims are anchored to the technical reports and to academic and university analyses (Liu, 2025; Mok, 2025; Woods, 2025). Benchmark scores are vendor-reported and had not been independently reproduced for the 0813 build as of this writing.
Appendix A: Glossary
The following entries define the technical vocabulary used in this article in language intended for an educated reader who is not a specialist in machine learning. Entries appear in alphabetical order.
Active parameters. The subset of a model’s parameters that actually run for a given input. In a mixture-of-experts model, only the parameters of the chosen experts are active; the rest remain idle until a different input routes to them.
Agentic (task). A task in which the model acts over multiple steps on the user’s behalf — calling tools, writing and running code, and carrying out multi-step workflows — rather than answering a single question.
Attention. The mechanism by which a transformer decides which parts of its input to emphasize when producing each output; a “spotlight” that shines on the most relevant words.
Cache hit / cache miss. In API billing, input text the model has already seen and stored (a cache hit) costs less than new text (a cache miss).
Chain-of-thought. Prompting or training a model to work through a problem step by step before giving its final answer, which improves performance on multi-step reasoning.
Context window. The amount of text a model can hold in view at once — effectively its working memory. The V4 models have a one-million-token window.
Distillation. Training a smaller “student” model to imitate a larger “teacher” model, transferring much of the larger model’s capability at a fraction of the cost.
FP8. An 8-bit floating-point number format — the specific low-precision format DeepSeek uses to cut the memory and compute required for training. See Low-precision training.
Group-relative policy optimization (GRPO). DeepSeek’s reinforcement-learning method for teaching models to reason; it scores a model’s answer by comparing it against a group of the model’s own alternative answers.
Inference. Running a trained model to produce answers. Inference is what users are billed for, in contrast with training, which builds the model.
Key-value (KV) cache. The stored intermediate results a model keeps for text it has already processed, so it does not recompute them. Compressing this cache is what makes long inputs affordable.
Low-precision training. Storing and computing a model’s numbers in fewer bits than usual so that training uses less memory and runs faster. Neural networks represent their parameters as floating-point numbers; standard training uses 16- or 32-bit precision, while low-precision training uses 8 bits. Fewer bits per number means the model occupies less memory and the hardware can process more numbers per second, at the cost of small rounding errors that can slightly reduce accuracy. The practical advance has been finding ways to keep that accuracy loss negligible — for example, by mixing higher- and lower-precision calculations — so the speed and memory savings come nearly free. It is roughly analogous to doing arithmetic to two decimal places instead of eight: faster and cheaper, and close enough for the purpose.
Mixture-of-experts (MoE). An architecture that divides a model’s capacity among many specialized sub-networks (“experts”) and uses a router to activate only a few of them per input — like a hospital triage nurse sending each patient to the relevant specialists.
Multi-head latent attention (MLA). DeepSeek’s attention variant that compresses the key-value cache into a smaller “latent” representation, cutting the memory needed to serve long contexts.
Open weights. Model parameters published for free download (here under the MIT license), allowing others to run, modify, and fine-tune the model. Contrast with proprietary models, whose weights are kept secret.
Parameters. The adjustable values inside a neural network that are tuned during training — the “knobs” that encode what the model has learned. Total parameters measure capacity; active parameters measure what actually runs per token.
Proximal policy optimization (PPO). A widely used reinforcement-learning algorithm that updates a model’s behavior in small, stable steps; the foundation that GRPO extends.
Quantitative hedge fund. An investment fund whose trading decisions are made by mathematical models, statistics, and computer algorithms rather than human judgment. Such funds typically employ machine learning and substantial computing resources; DeepSeek’s parent company, High-Flyer, is a Chinese quantitative hedge fund.
Reasoning effort. A user-selectable setting (low, high, or max) that controls how much computation a model spends “thinking” before answering; higher effort suits harder tasks.
Reinforcement learning (RL). Training a model by rewarding or penalizing its outputs so it improves through trial and error, rather than by imitating examples alone.
Router. In a mixture-of-experts model, the small network that decides which experts process each token.
Scaling laws. Empirically observed relationships between model size, training data, compute, and performance, used to predict how much of each resource is needed.
Sparse activation. The property of running only a small fraction of a model’s parameters for any single input — about 3% for V4-Pro — so that capacity can grow without a proportional rise in cost.
Token. A small chunk of text (roughly a word or part of one) that a model processes as a unit. Prices and context lengths are measured in tokens.
Transformer. The dominant neural-network architecture for language models, introduced in 2017 and built around attention mechanisms.
Training. The process of building a model by exposing it to vast amounts of text and adjusting its parameters; contrast with inference.
Appendix B: DeepSeek’s Chinese Name — Translation and Meaning
The company’s registered name, 杭州深度求索人工智能基础技术研究有限公司, translates as “Hangzhou DeepSeek Artificial Intelligence Basic Technology Research Co., Ltd.” Chinese company names follow a fixed sequence — place, brand name, business scope, and legal form — and the components of this one map onto that pattern as follows:
杭州 (Hángzhōu) — Hangzhou, the city where the company is registered (capital of Zhejiang province).
深度求索 (shēndù qiúsuǒ) — “deep seeking,” the brand name from which the English “DeepSeek” is translated. 深度 (shēndù) means “deep” (depth); 求索 (qiúsuǒ) means “to seek, to search, to explore.”
人工智能 (réngōng zhìnéng) — “artificial intelligence.” 人工 (réngōng) is “man-made” (人 = person/human; 工 = work/labor); 智能 (zhìnéng) is “intelligence.”
基础技术 (jīchǔ jìshù) — “basic/fundamental technology” (基础 = foundation; 技术 = technology).
研究 (yánjiū) — “research.”
有限公司 (yǒuxiàn gōngsī) — “limited company” (有限 = limited; 公司 = company), the standard corporate suffix equivalent to “Co., Ltd.”
Two details in the name carry meaning beyond the literal. First, the word 求索 (“to seek”) is drawn from one of the most famous lines in classical Chinese poetry — Qu Yuan’s “Li Sao” (c. third century BCE), 路漫漫其修远兮，吾将上下而求索 (“The road ahead is long and far; I shall search high and low”). The “seek” in DeepSeek therefore carries a scholarly, aspirational sense of an open-ended quest for knowledge rather than a commercial “search,” which fits the company’s self-presentation as a research lab pursuing artificial general intelligence. Second, the registered scope is “basic technology research” (基础技术研究), signaling that the entity was established as a research organization rather than a product or software vendor — consistent with its origin as a lab spun out of the High-Flyer hedge fund.
A reader might notice that 人工, the two characters meaning “artificial” or “man-made,” faintly resemble the letters “AI.” This is a coincidence, not an etymology. Chinese characters are logograms — each stands for a meaning (and usually a sound) rather than a letter — and their shapes descend from ancient drawings. 人 began as a picture of a person and now resembles an inverted V (∧) rather than an “A,” while 工 began as a picture of a carpenter’s tool and genuinely resembles a capital “I.” The resemblance is therefore accidental — though a fitting one, since the phrase 人工智能 maps cleanly onto the English “artificial intelligence,” and the abbreviation “AI” is borrowed into Chinese directly from the English initials.
Note on the Research and Writing Process
This article was researched and drafted in collaboration with Hermes, an AI agent developed by Nous Research, which in this engagement operated on the deepseek-v4-pro model. Hermes contributed a set of specialized research skills — reusable workflows that govern how it locates, verifies, and formats sources — together with live web search and full-text retrieval across primary materials (DeepSeek’s release notes and pricing documentation, Reuters, and Quartz); verification of every reference’s DOI, author list, and publication date against the Crossref and DataCite registries; extraction of the benchmark and pricing tables directly from DeepSeek’s release materials; and an APA 7 authoring protocol governing in-text citations, the reference list, tables, and the appendices. I served as research lead throughout: I framed the research question and its scope, set the source standards (APA 7, DOI-verified citations, and a preference for peer-reviewed and foundational research over commentary on the product itself), and directed the piece through substantive revision — reframing the peer-reviewed dimension as the evidence-based knowledge underlying DeepSeek’s methods, requesting the plain-language explanations, analogies, and glossary, and requiring a full accuracy review before publication. The division of labor illustrates a productive human–AI partnership for research writing: the agent supplied speed, breadth of retrieval, and citation and formatting rigor, while the analytical framing, editorial judgment, and domain knowledge remained mine.