# DeepSeek hiring roles · skill map and full JDs

> Interactive page (recommended): [DeepSeek招聘职位技能分析.html](../../调查报告/DeepSeek招聘职位技能分析.html)
>
> Source: [talent.deepseek.com](https://talent.deepseek.com/) · site snapshot 2026-09-07T06:09:32Z · **34** open roles in total
>
> JD body text is kept **as in the official HTML**. Skill tags are extracted from the body by rules, for filtering and counts; they are not official tags. Apply via the official site / Moka.

The official site groups roles as “full-stack development / algorithms, AI core-systems R&D, operations, product, model-data strategy, deep-learning researcher, functions.” That cut is useful for applying, but too coarse for “what should I practice”—“full-stack development / algorithms” simultaneously holds front-end, server-side, Agent Harness, pre-training data, and testing. This page recuts by **work content** into 10 tracks, and counts in how many roles a skill appears (each role counted at most once; required = appears before “plus / nice-to-have”).

---

## How to use the interactive page

White-background reader: click a role name on the left; read the full JD on the right.

1. **Category / location** is a Chip row, selected by horizontal scroll; skills stack via a dropdown. The current filter shows as removable tags.
2. Each JD shows at most **6 skill Chips** above the body; the rest sit under “more skills.” Clicking a Chip filters the list by that skill.
3. Multi-track roles (for example server-side) use direction Tabs; **only one direction is expanded at a time**.
4. “Highlight skills” is off by default, so the original text stays clean; when on, skill terms get a blue wash (no 【】, and cultural words such as open source are not marked). The switch is written into the address bar as `?hl=1`.
5. The top-bar **skill notes** are 5 short cards + the first 8 skill bars. Clicking a bar or a category returns to the role list with that filter applied.
6. Arrow keys or `j` / `k` move up and down in the list. On a phone, tap the list first, then “← role list” to go back.

---

## Track mapping (headcount)

| Track | Roles | Rough official-site home |
|---|---:|---|
| Functional support | 7 | HR / legal / finance / procurement / admin / management trainee / cross-domain |
| Data engineering / data strategy | 6 | Pre-training data + model-data strategy |
| Deep-learning research | 5 | Researchers + AI search |
| Testing / operations / infrastructure | 4 | Operations + test development |
| LLM training / inference | 3 | Framework / operator compilation / deep-learning R&D |
| Supercomputing / storage / clusters | 3 | AI core-systems R&D |
| Agent engineering | 2 | Full-stack development / algorithms |
| Product / operations | 2 | Product department |
| Front-end / client | 1 | Full-stack development / algorithms |
| Back-end / server-side | 1 | Full-stack development / algorithms (contains 6 directions) |

Locations counted by role (one role may list multiple cities): **Beijing 28 · Hangzhou 30 · Ulanqab 2** (IDC, procurement).

---

## Skill occurrence counts (every tag that hits)

| Rank | Skill | Roles | Required | Plus | Group |
|---:|---|---:|---:|---:|---|
| 1 | Cross-team collaboration | 20 | 17 | 3 | Collaboration |
| 2 | Open-source contribution | 15 | 6 | 9 | Evidence |
| 3 | Inference systems / KV Cache | 14 | 10 | 4 | Model training |
| 4 | Agent / Harness | 11 | 11 | 0 | Agent |
| 5 | Python | 11 | 9 | 2 | Language |
| 6 | Rust | 11 | 9 | 2 | Language |
| 7 | C++ | 10 | 9 | 1 | Language |
| 8 | Storage / filesystems | 10 | 9 | 1 | Systems engineering |
| 9 | Data pipelines / cleaning | 9 | 8 | 1 | Data |
| 10 | Evaluation / Benchmark | 9 | 8 | 1 | Model training |
| 11 | Linux / kernel / OS | 9 | 5 | 4 | Systems engineering |
| 12 | Distributed systems | 8 | 6 | 2 | Systems engineering |
| 13 | Papers / top conferences | 8 | 1 | 7 | Evidence |
| 14 | GPU / NPU / 10k-GPU clusters | 7 | 6 | 1 | Model training |
| 15 | High-performance networking / RDMA | 7 | 4 | 3 | Systems engineering |
| 16 | Contests (NOI/ICPC, etc.) | 7 | 0 | 7 | Evidence |
| 17 | Product sense / user insight | 6 | 6 | 0 | Product / org |
| 18 | Coding with agents | 6 | 5 | 1 | Agent |
| 19 | Post-training / RL | 6 | 4 | 2 | Model training |
| 20 | High concurrency / high availability | 6 | 3 | 3 | Systems engineering |
| 21 | PyTorch | 6 | 2 | 4 | Model training |
| 22 | Multimodality | 5 | 5 | 0 | Model training |
| 23 | Go / Golang | 5 | 4 | 1 | Language |
| 24 | Observability / monitoring | 5 | 4 | 1 | Systems engineering |
| 25 | Databases / caches / queues | 5 | 3 | 2 | Systems engineering |

The fuller set of 45 items is in the interactive page’s tag area. The conclusions below are a comparison reading, not official-site wording.

---

## Comparison conclusions

**1. This is not an internet-company org chart of “one front-end team + one back-end team.”**  
There is only 1 front-end role (web + two directions, iOS/Android). There is 1 server-side role, but it splits internally into six directions: research platform, Agent framework, R&D-efficiency infrastructure, API, online services, data engineering. The real headcount density is in **data strategy, researchers, training / supercomputing systems**.

**2. Language stack: Python and Rust are tied (11 roles each), C++ 10 roles.**  
Systems roles almost always list Rust + C++ together, not as an either/or. Go appears in server-side / testing / operations development. JavaScript/TypeScript is a required skill only on the front-end role. If you want to match DeepSeek technical roles, the priority order is closer to `Python + (Rust|C++) + Linux/distributed` than to the front-end ecosystem.

**3. Agent has moved from a product word to a hiring baseline.**  
11 roles mention Agent / Harness directly in the body, and all of those count as required—none appear only under plus. Server-side, Harness, product, design, and data product manager also require “you yourself are a heavy user” or the ability to write code with an agent. The Harness role even names Claude Code / Codex / Cursor / OpenCode / Manus / Hermes.

**4. “LLM training” is a vertically sliced system, not a job title.**

| Layer | Roles | More specific required skills |
|---|---|---|
| Data fuel | Pre-training data engineer, Code Agent data engineer, data product manager | Pipelines, evaluation, domain taste, RL environments |
| Algorithms | Pre-training / post-training / multimodal / Frontier researchers | Training dynamics, RLHF/GRPO, continual learning, vision encoders |
| Framework | LLM training / inference framework engineer, deep-learning R&D engineer | PyTorch, MoE, long context, asynchronous RL, KV Cache |
| Operators | High-performance operator / communication / compiler | CUDA / Ascend C / TileLang, NCCL, RDMA |
| Cluster and storage | Supercomputing, distributed storage, R&D infrastructure | Scheduling, RDMA, filesystems, cross-layer debugging |
| Treating the cluster as utilities | AI platform operations, IT, IDC | Linux, GPU-cluster stability, liquid cooling / power (IDC) |

**5. Open source is a cultural signal.**  
15 roles mention open source; 9 of those place it under plus. The R&D infrastructure role states that they look at GitHub **commit history**, not star counts. Papers and NOI/ICPC are almost entirely plus (required almost never), which means they are plus leverage, not an entry ticket.

**6. One JD is empty.**  
“AI product operations (experience and service)” has official-site body text of only “（新部门，虚位以待）” (“new department, position held open”). The interactive page keeps it as-is.

---

## All roles (by track)

### Front-end / client (1)

- **Front-end / client development engineer** · full-stack development / algorithms · Beijing / Hangzhou · intern · full-time

### Back-end / server-side (1)

- **Server-side development engineer** · full-stack development / algorithms · Beijing / Hangzhou · intern · full-time

### Agent engineering (2)

- **Agent elastic-compute R&D engineer** · full-stack development / algorithms · Beijing / Hangzhou · intern · full-time
- **Agent Harness team** · full-stack development / algorithms · Beijing / Hangzhou · intern · full-time (researcher / R&D / DevRel / product / design / project manager / operations)

### LLM training / inference (3)

- **Deep-learning R&D engineer** · full-stack development / algorithms · Beijing · intern · full-time
- **High-performance operator / communication / compiler engineer** · AI core-systems R&D · Beijing / Hangzhou · intern · full-time
- **LLM training / inference framework engineer** · AI core-systems R&D · Beijing / Hangzhou · intern · full-time

### Supercomputing / storage / clusters (3)

- **Supercomputing-cluster R&D engineer** · AI core-systems R&D · Beijing / Hangzhou · intern · full-time
- **R&D infrastructure engineer** · AI core-systems R&D · Beijing / Hangzhou
- **High-performance distributed-storage engineer** · AI core-systems R&D · Hangzhou · intern · full-time

### Deep-learning research (5)

- **AI search algorithm / architecture engineer** · full-stack development / algorithms · Beijing · intern · full-time
- **Frontier (continual learning / self-evolution / new paradigms) researcher** · deep-learning researcher · Beijing · intern · full-time
- **Pre-training (data / algorithms) researcher** · deep-learning researcher · Beijing · intern · full-time
- **Post-training (data / algorithms) researcher** · deep-learning researcher · Beijing / Hangzhou · intern · full-time
- **Multimodal understanding (data / algorithms) researcher** · deep-learning researcher · Beijing / Hangzhou · intern · full-time

### Data engineering / data strategy (6)

- **Pre-training data engineer** · full-stack development / algorithms · Beijing / Hangzhou · intern · full-time
- **Code Agent data engineer** · model-data strategy · Beijing · intern · full-time (front-end / mobile / back-end / Web3D / games / security specialty)
- **General-agent data product manager (office / daily life / search)** · model-data strategy · Beijing / Hangzhou · intern · full-time
- **Specialized-domain data product manager (low-resource languages, medicine, law, and other disciplines)** · model-data strategy · Beijing / Hangzhou · intern · full-time
- **AI creation data product manager** · model-data strategy · Beijing / Hangzhou · intern · full-time
- **Affective-intelligence data product manager** · model-data strategy · Beijing / Hangzhou · intern · full-time

### Testing / operations / infrastructure (4)

- **Test-development engineer** · full-stack development / algorithms · Hangzhou · intern · full-time
- **AI-platform operations engineer** · operations · Hangzhou · intern · full-time
- **IT infrastructure team** · operations · Beijing / Hangzhou · intern · full-time
- **IDC datacenter team** · operations · Beijing / Ulanqab / Hangzhou · full-time

### Product / operations (2)

- **AI product manager** · product department · Beijing / Hangzhou · intern · full-time
- **AI product operations (experience and service)** · product department · Beijing / Hangzhou · JD original text almost empty

### Functional support (7)

- **AI cross-domain technical talent** · full-stack development / algorithms · Beijing / Hangzhou · intern · full-time
- **HR team** · Beijing / Hangzhou · full-time
- **AGI core-business management trainee** · Beijing / Hangzhou · intern · full-time
- **Legal team** · Hangzhou · full-time
- **Finance team** · Hangzhou · full-time
- **Procurement team** · Ulanqab / Hangzhou · full-time
- **Administration team** · Beijing / Hangzhou · intern · full-time

Full responsibilities, requirements, and plus items are in the interactive page’s right-hand detail pane, or in each role’s Moka link.

---

## Source data and regeneration

- `调查报告/deepseek-talent-jobs.json`: original text of the 34 roles extracted from the official site’s JS bundle via `JSON.parse`
- `scripts/build-deepseek-talent-html.py`: assigns tracks, extracts skills, generates the interactive HTML

```bash
python3 scripts/build-deepseek-talent-html.py
```