DeepSeek upgrades chatbot with advanced 'thinking' feature

DeepSeek, an AI startup based in Hangzhou, has added an “interleaved thinking” feature to its chatbot, allowing it to perform multi-step research tasks.

The new function lets the chatbot process information in stages, showing its reasoning at each step, and is available on both DeepSeek’s website and mobile app.

This upgrade follows a reported 90% rise in DeepSeek’s monthly active users in December, reaching nearly 131.5 million, according to AI product tracker Aicpb.com.

The “interleaved thinking” capability was introduced in DeepSeek’s latest V3.2 model, released in early December.

DeepSeek’s latest updates also include a timeline interface that lets users switch between questions, a feature seen in few other chatbots.

Industry watchers expect DeepSeek to release a new major model around the upcoming Spring Festival in February.

🔗 Source: South China Morning Post

🧠 Food for thought

Implications, context, and why it matters.

Independent checks on interleaved thinking remain limited

DeepSeek V3.2 adds DeepSeek Sparse Attention (DSA) 1, which uses a learned “lightning indexer” (a selector that flags the most relevant parts of the input) to focus on relevant tokens (the chunks of text AI models process) and cut attention complexity from quadratic to linear 1. Interleaved thinking ships with V3.2, yet materials do not tie it to DSA.
DeepSeek cites “minimal impact on output quality” for V3.2-Exp 2. No third-party benchmarks confirm multi-step research accuracy versus Moonshot AI or MiniMax. Training also borrows from DeepSeekMath V2 with separate verifier models to boost proof quality 1, though outside validation is missing.
Monthly active users grew 90% to 131.5 million, yet standardized tests remain scarce.

API access to interleaved reasoning could enable research-agent workflows

Pay-as-you-go pricing lists $0.56 per million input tokens for cache misses, $1.68 per million output tokens, and $0.07 for cache-hit input 3.
deepseek-chat and deepseek-reasoner run on DeepSeek-V3.2 4. API access to interleaved thinking is unclear, which suggests it might be limited to the website and mobile app.
Builders need endpoint specs, rate limits and access terms to design at-scale agents. If exposed, teams can maximize cache hits at $0.07 per million tokens versus $0.56 for misses 3. Reused prompts plus a 128,000-token context window (the amount of text the model can consider at once) can cut costs by up to 75% 3.