OPEN-WEIGHT · REASONING
DeepSeek
Updated 4 months agoChinese open-weight model family with strong code + math benchmarks.
- Maker
- DeepSeek
- Since
- 2023
- From
- Free chat
OPEN-WEIGHT · REASONING
Chinese open-weight model family with strong code + math benchmarks.
About DeepSeek
DeepSeek is a large‑language‑model family developed by the Chinese research lab DeepSeek and first released in November 2024. The models are positioned as open, efficient reasoning systems that aim to offer strong price‑to‑performance characteristics within the broader LLM landscape.
DeepSeek Near $7 Billion AI Funding Deal
DeepSeek is close to finalizing a $7.4 billion funding deal, one of the largest startup financings in China. The investment includes Tencent Holdings Ltd. and the company's founder. This deal highlights the growing strength of China's domestic AI industry, which is seeking capital to compete with major U.S. players like OpenAI. The funding is expected to accelerate the development of AI technologies and strengthen the company's position in the global market.
Bloomberg — Technology
Code Shapes AI Agent Behavior
A new review paper suggests that the software layer surrounding a language model, rather than the model itself, is the key to creating autonomous AI agents. This 'harness' includes tools, memory, testing, and permission boundaries that enable stateless models to function as working agents. A dedicated team at Deepseek in Beijing is already exploring this concept, with a core formula that supports the thesis: model + harness = AI agent.
The Decoder
Posts to your status feed
Pick the closest match below, edit the body, and post. Your report carries the #deepseek tag automatically so it surfaces here + in the trending-tags rail. Reporting also follows DeepSeek so you’ll get status updates.
Free chat · MIT-licensed open weights · API ~$0.14/M input · $0.28/M output (V3)
Best-effort summary — confirm on the provider's own pricing page; tiers and prices drift.
DeepSeek Near $7 Billion AI Funding Deal
DeepSeek is close to finalizing a $7.4 billion funding deal, one of the largest startup financings in China. The investment includes Tencent Holdings Ltd. and the company's founder. This deal highlights the growing strength of China's domestic AI industry, which is seeking capital to compete with major U.S. players like OpenAI. The funding is expected to accelerate the development of AI technologies and strengthen the company's position in the global market.
Bloomberg — Technology
Code Shapes AI Agent Behavior
A new review paper suggests that the software layer surrounding a language model, rather than the model itself, is the key to creating autonomous AI agents. This 'harness' includes tools, memory, testing, and permission boundaries that enable stateless models to function as working agents. A dedicated team at Deepseek in Beijing is already exploring this concept, with a core formula that supports the thesis: model + harness = AI agent.
The Decoder
No articles yet in DeepSeek. Check back soon, or browse all sections.
No forum thread for DeepSeek yet — start one with the button above.
Loading edit history…
DeepSeek V4 Pro is the current flagship model in the DeepSeek family, released in April 2026 with a knowledge cutoff of January 2026. It comprises roughly 1.6 trillion parameters overall and uses a mixture‑of‑experts architecture with about 49 billion active experts, supports a context window of up to one million tokens and can generate outputs as long as 384 k tokens; it also introduces an optional “thinking mode” that subsumes the earlier R1 capability. The model is distributed under an MIT license and is priced at $0.435 for input and $0.87 per million output tokens.
DeepSeek V4 Flash is an active, fast‑tier model in the DeepSeek family released on 1 April 2026 with a knowledge cutoff of January 2026. It serves as a budget‑oriented, high‑volume variant of the V4 series, offering a 1 million‑token context window and operating under an MIT license, with pricing of $0.14–$0.28 per million tokens. The model comprises 284 billion parameters overall, of which 13 billion are active in its mixture‑of‑experts architecture.
DeepSeek‑V3 is a large language model in the DeepSeek family that was released on 26 December 2024 and features an expanded context window of up to 128 000 tokens. It is classified as a legacy, superseded tier within the series, indicating it has been succeeded by later versions.
DeepSeek‑R1 is a reasoning‑tier language model in the DeepSeek family that was released on 20 November 2024 and features an extended context window of up to 128 000 tokens. The model has since been superseded by later versions within the same series.
DeepSeek Coder V2 is an open‑weight, coding‑specialized language model released in June 2024 by the DeepSeek team under an MIT license. It builds on earlier DeepSeek Coder releases and is noted for strong performance on HumanEval benchmarks, making it a popular choice for local development environments. The model has since been superseded by newer versions.
DeepSeek V2 is a version of the DeepSeek language model series that was released on 1 May 2024. It belongs to the deepseek family and is classified as a legacy‑tier model that has since been superseded by newer releases. Specific changes relative to earlier DeepSeek versions are not detailed in the available data.
DeepSeek V1 is the inaugural large language model released by deepseek on November 1 2023 and is now classified as a legacy, superseded version. As the company’s first LLM, it was positioned as competitive with larger contemporary models and served to introduce the DeepSeek brand to international AI communities.
DeepSeek V4 promotional pricing made permanent
On 2026-05-22 DeepSeek made its 75% promotional discount the permanent list price: V4 Pro $0.435/$0.87, V4 Flash $0.14/$0.28 per 1M.
DeepSeek V4 Pro released
Flagship; 1.6T/49B MoE, 1M context, thinking mode; launched at promotional 75% pricing.
DeepSeek V4 Flash released
Budget sibling; 284B/13B MoE, 1M context; launched at promotional 75% pricing.
DeepSeek R1 released
Reasoning model competitive with OpenAI o1 at ~95% lower cost; the industry 'Sputnik moment'.
DeepSeek V3 released
Catch Me Up — everything that changed across AI & smart home ↗
685B/37B trained on 14.8T tokens; matched GPT-4o on most benchmarks at a fraction of training cost.
DeepSeek Coder V2 released
Coding-specialized MIT-licensed model with strong HumanEval performance.
DeepSeek V2 released
MLA + DeepSeekMoE architecture; 236B/21B; $0.14/M input launched the Chinese AI price war.
DeepSeek V1 released
DeepSeek's entry into the LLM market, competitive against much larger models.