The engineers behind DeepSeek and the shift toward open-weight reasoning

DeepSeek researchers utilize reinforcement learning and the GRPO algorithm to develop reasoning models like DeepSeek R1. These open-weight models achieve a 49.2% SWE-bench accuracy and have significantly disrupted the market, causing an 18% drop in Nvidia’s share price.

The engineers behind DeepSeek and the shift toward open-weight reasoning

Engineering reasoning through reinforcement learning

Liang Wenfeng manages DeepSeek with a research culture that avoids mandatory overtime and rigid performance indicators. He founded the company in 2023 after spinning off a research lab from the High-Flyer hedge fund in July 2023. DeepSeek researchers, often recruited from Peking, Tsinghua, and Beihang universities, focus on improving model capabilities through reinforcement learning. This training method incentivizes the development of reasoning patterns such as self-reflection and verification without needing human-labeled reasoning trajectories. DeepSeek R1 uses the DeepSeek V3 architecture, which incorporates Mixture-of-Experts and Multi-Head Latent Attention. This architecture compresses key and value tensors into a lower-dimensional space before storage in the KV cache. DeepSeek operates Fire-Flyer 2, a computing cluster with 5,000 PCIe A100 GPUs in 625 nodes. The cluster uses the 3FS distributed parallel file system and the hfreduce library for asynchronous communication. The platform supports cross-zone tasks through a network topology of two fat trees. DeepSeek R1 achieves superior performance on verifiable tasks like mathematics and coding competitions by using reinforcement learning to facilitate the development of advanced reasoning patterns such as self-reflection, verification, and dynamic strategy adaptation.

Performance comparison and cost advantages

DeepSeek R1 provides detailed, step-by-step explanations for its answers, which aids debugging and learning in technical fields. While ChatGPT o3-mini returns concise code optimizations, DeepSeek R1 explains why it made every specific change to a function. I find the 32 second response time for DeepSeek R1 compared to 6 seconds for o3-mini frustrating when you need fast results. DeepSeek R1 holds a 1900 Codeforces rating and a 49.2% SWE-bench accuracy. In a task for 3D animation generation using Python, DeepSeek R1 produced functional output while ChatGPT o3-mini failed to deliver. For a math problem regarding the sum of squares of the first 10 prime numbers, DeepSeek R1 breaks down each prime number and its square to provide a full explanation.

Metric DeepSeek R1 ChatGPT o3-mini
Codeforces Rating 1900 2130
SWE-bench Accuracy 49.2% Not specified
Input Token (Cache Hit) $0.14 / 1M Not specified
Input Token (Cache Miss) $0.55 / 1M $1.10 / 1M
Output Token $2.19 / 1M $4.40 / 1M

You likely know that model pricing dictates which tools a developer selects for a production pipeline. DeepSeek R1 provides input tokens for $0.14 per million during a cache hit. This price is much lower than the $1.10 minimum input cost for OpenAI’s o3-mini. DeepSeek R1 also manages context better than o3-mini, which missed a subtle detail in a contextual logic test. While o3-mini generates more tokens, DeepSeek R1 produces concise and focused outputs. The model uses the GRPO algorithm, a variant of Proximal Policy Optimization, to improve reasoning via Reinforcement Learning with Verifiable Rewards. Recent releases like DeepSeek V3.1 and V3.2 shifted towards hybrid models where users switch between reasoning and non-reasoning modes via chat prompts.

Global adoption and market disruption

The release of DeepSeek-R1 in January 2025 triggered an 18% drop in Nvidia’s share price. This event removed approximately $108 billion from the wealth of 500 billionaires. DeepSeek’s open-weight models, released under the MIT license, allow developers to download and fine-tune reasoning paths directly. This accessibility drives adoption in Southeast Asia and Africa, where companies in Nairobi use the tools for local language models. In April 2026, the company released the V4 series, including the 1.6-trillion parameter DeepSeek-V4-Pro. The V4.1-Flash model, released in September 2026, uses a causal encoder-decoder architecture and consumes less memory than previous versions. DeepSeek’s popularity also reached the US, where it surpassed ChatGPT as the most downloaded freeware app on the iOS App Store in late January 2025. The V3.2-Exp model, which used V3.1-Terminus as its base, introduced DeepSeek Sparse Attention, which uses a lightning indexer and a token-selector to improve efficiency. This technology allows the model to selectively reduce context for better performance in long-context scenarios. Semiconductor manufacturers such as Huawei and Cambricon Technologies have adopted the V4 series. The release of R1 in January 2025 provided a blueprint for the industry, showing that rapid progress is possible with limited resources. This led to an explosion of Chinese open-source releases, with Baidu increasing its Hugging Face releases from zero in 2024 to over 100 in 2025. Will the success of these open-weight models eventually force US companies to abandon proprietary APIs?

airtrain.ai
airtrain.ai

The airtrain.ai newsroom covers AI research, models and the tools built on them.

More on this topic

Stay ahead of AI

Get the week's most important AI stories delivered to your inbox every Monday.

No spam. Unsubscribe anytime.

More Stories