AI tools·Chat models·China
DeepSeek
China’s flagship open-weight + API lab. R1 put it on the global map; still extreme value.
Compiled through 2026-09
Open weights, self-hostable
- Vendor
- DeepSeek
- Started
- 2023
- China availability
- Works in ChinaDirectly reachable from mainland China
- Pricing tier
- Free tier
- Current flagship
- DeepSeek-V3 / DeepSeek-R1
- Website
- www.deepseek.com
Evaluation, Pros & Cons
Editorial score based on public benchmarks and community feedback — not a third-party benchmark result
Key Advantages
- Frontier open-weights pioneer; V3/R1 rival proprietary flagships at fractional cost
- R1 shows breathtaking mathematical and coding depth with native Chinese fluency
Limitations & Caveats
- Official web portal and API experience occasional queuing during global traffic spikes
Overview
DeepSeek is known for MoE efficiency. V3 (late 2024) and R1 (Jan 2025, MIT) matched then-o1-class reasoning at a fraction of the cost. V3.1/V3.2 unified chat and thinking; 2026’s V4 (Flash / Pro) pushed context and agents. Web, app, and an OpenAI-compatible API make it a default base model in China.
Direction
The public story remains open weights, very low inference cost, and stronger code/math/agents. A next reasoner (R2 was discussed through 2025) sits alongside V4’s tool-use and terminal focus. Best as a cheap, self-hostable base—not always the best writer.
Version history
May 2026
LatestDeepSeek V4 Flash / Pro
Next flagship; longer context and agents.
View details
New capabilities
- V4 Flash / Pro split
- Longer context and terminal/agent engineering
What got stronger
- More plausible as a business base, not only cheap chat
Version notes
V4 pushed DeepSeek from viral value to production agents. Flash for volume, Pro for hard jobs. Writing style still may trail Claude; engineering and price are the home field.
Jun 2025
DeepSeek-V3.1 / V3.2
Hybrid thinking; better tool use.
View details
New capabilities
- Hybrid chat + thinking modes
- Better tool use, JSON, and agents
What got stronger
- No separate model just to “think”
Version notes
V3.1 / V3.2 folded R1-style thinking into the daily chat model and fixed tool use. For app builders that mattered more than another contest score.
Jan 2025
DeepSeek-R1
Reasoning model, MIT license, o1-class.
View details
New capabilities
- R1 reasoner, MIT license
- o1-class contest/math performance at the time
What got stronger
- Open reasoning that the world could run locally or cheaply
Version notes
R1 was the early-2025 shock: reasoning left the closed-only club. Distills followed so consumer GPUs could follow. Hallucinations and safety policy remain on you.
Dec 2024
DeepSeek-V3
671B MoE open weights; breakout value.
View details
New capabilities
- 671B MoE open weights
- OpenAI-compatible API at a very low price
What got stronger
- Forced the industry to rethink inference pricing
Version notes
V3 showed a China lab could reach near-frontier closed quality with MoE and then release the weights. It changed “is this worth the price?”, not only leaderboards.