Why this lesson exists
You are about to hand real work to a model built by a company most Western developers had never heard of eighteen months ago. Understanding why it exists explains almost everything about how it behaves — its bluntness, its pricing, and its willingness to show you its thinking.
A quant fund that got curious
DeepSeek did not begin as a startup chasing a consumer chatbot. It began inside High-Flyer, a Chinese quantitative hedge fund that had spent years buying and running GPUs to find signals in financial markets. That origin matters. A quant fund already owned the two things frontier model research demands: large-scale compute and people who think in terms of measured, compounding efficiency rather than headline spending.
DeepSeek was spun out as its own research lab in 2023, led by High-Flyer co-founder Liang Wenfeng, and funded by the fund rather than by venture capital. That structure bought it something unusual: no investor pressure for a fast consumer product, and a mandate to publish.
The consequence: DeepSeek releases open weights under permissive licences. You can download the model, run it on your own hardware, read the architecture, and audit the technical report. Very few frontier-scale labs do that.
The four releases that matter
You can understand the whole company through four moments.
| Release | What it introduced | Why a builder should care |
|---|---|---|
| DeepSeek-V2 May 2024 |
Multi-head latent attention (MLA) | Collapsed the memory cost of long conversations. This is the technical root of everything that makes DeepSeek cheap to run as an agent today. |
| DeepSeek-V3 December 2024 |
A 671-billion-parameter mixture-of-experts model | Demonstrated that a genuinely frontier-class model could be trained for a fraction of the assumed cost — and be given away. |
| DeepSeek-R1 January 2025 |
Reasoning learned almost entirely through reinforcement learning | Proved reasoning behaviour could emerge from RL without supervised examples of reasoning. It also moved public markets. |
| DeepSeek-V4 / V4.1-Flash 2026 |
A new architecture family with native multimodal vision | The generation this course targets. V4.1-Flash is the smallest model in the family, and the one you will actually run. |
The R1 moment, and what it actually changed
Before R1, the received wisdom was that teaching a model to reason step-by-step required enormous quantities of human-written reasoning traces. R1-Zero demonstrated something stranger: give a capable base model a verifiable reward signal and enough reinforcement learning, and structured reasoning behaviour emerges — including the model spontaneously producing long chains of thought, checking its own work, and rethinking dead ends.
The release landed in late January 2025 and triggered a violent reassessment in public markets of what frontier AI should cost to produce. That part of the story is well covered in the press and you can skip most of it.
What matters for you as a practitioner is narrower and more durable:
- Reasoning became visible. R1-family models emit thinking tokens you can read. You can see the model plan, wander, correct itself, and commit. That is an enormous debugging affordance, and this course leans on it heavily.
- Reasoning became controllable. Because it is a trained behaviour rather than a hidden process, it can be dialled up and down per request. Lesson 9 covers the effort levels.
- The cost floor moved. Open weights at frontier quality meant the price of inference stopped being set by whoever owned the best closed model.
The misquoted number
You will see "$5.6 million" cited constantly as the cost of training DeepSeek-V3. Read the V3 technical report and you find that figure describes the GPU rental cost of the final training run — it excludes research, failed runs and ablations, staff, and the hardware itself. It is a real and striking number. It is not "the cost of building DeepSeek." Being precise about this distinction is a good habit to carry into every benchmark table you read in Module 1.
What "open-weight" buys you
The phrase gets used loosely, so pin it down. Open weights means the trained parameters are published under a licence you can read — DeepSeek's are permissive (MIT on the current models). You may download them, run them, fine-tune them, and build products on them. What is not published is the training data and the full training pipeline.
Three practical consequences:
- You can self-host. For regulated work, or when a task is high-volume enough that per-token pricing is the wrong shape, running locally is a real option. Weigh it honestly — the headline model is enormous, so this is usually a cost-per-throughput question, not a "runs on a laptop" question.
- You are not locked to one vendor's API. Because the architecture is public, many inference providers serve it, and prices compete. Lesson 34 turns this into a routing strategy.
- You can verify claims. When someone says a model does something, the architecture and technical reports are checkable. You are not reading marketing.
The 2026 position
By the September 2026 architecture refresh, DeepSeek's public framing had shifted from the upstart to the incumbent-with-leverage. The V4.1-Flash release notes describe the new architecture family as designed "for a higher capability ceiling, faster inference, higher throughput, and scaling to larger models" — the language of a lab that expects to keep going, not one that got lucky once.
The release itself is deliberately unglamorous. V4.1-Flash is pitched as the smallest model in its family, with native visual understanding, released alongside a price reduction. That framing tells you where the company thinks the market is: not chasing the single most capable model, but making a very good model cheap enough to run constantly.
Checkpoint
- Which single architectural idea from 2024 is the reason long agent sessions are affordable today?
- What did R1-Zero establish that changed how people thought about training reasoning?
- State precisely what the $5.6M figure does and does not include.
- Name two things open weights let you do that a closed API does not.
Answers
- Multi-head latent attention (MLA), introduced with V2 — it cut the memory cost of long contexts.
- That reasoning behaviour can emerge from reinforcement learning against a verifiable reward, without supervised reasoning traces.
- It is the GPU rental cost of the final V3 training run. It excludes research, failed runs, staff, and hardware.
- Self-host for cost or compliance; verify architecture and behaviour against public reports; and avoid single-vendor lock-in.
Exercise: one paragraph, no jargon
Write four sentences — no more — explaining DeepSeek to a smart colleague who has never heard of it. Rules: no benchmark numbers, no parameter counts, and no mention of "open source" as a virtue in itself. You must say what it is, where it came from, why it got attention, and what it means for the tools they use.
Keep it. In Lesson 4 you will compare it against what you know then, and you will probably want to rewrite it. That gap is the point of the exercise.
Next
You now know where the model came from. Lesson 2 deals with the more practical question: which model string do you actually send, and what happens when you send the wrong one.