DeepSeek-V3.2 In-Depth Analysis: One Pragmatic, One Reaching for the Top

On the evening of December 1, 2025, DeepSeek released two official models at once: DeepSeek-V3.2 and DeepSeek-V3.2-Speciale. This "dual-model" strategy is clear—one aimed at everyday deployment, the other used to probe the ceiling of reasoning.

I. V3.2: The Everyday Workhorse with GPT-5-Level Performance

V3.2 is the official successor to the late-September experimental version V3.2-Exp and is now fully available on the official website, App, and API. Its positioning is to "balance reasoning capability and output length": it retains strong reasoning without letting token consumption spiral out of control, making it suitable for everyday scenarios such as Q&A, writing, and coding.

The official claim is "GPT-5 level performance," and its token usage is noticeably more economical—for the same problem, V3.2 Thinking's reasoning length is generally in the 16,000–20,000 token range, while some competitors need to run past 30,000.

II. V3.2-Speciale: The Lopsided Research Genius

Speciale is the long-thinking enhanced version of V3.2, incorporating the theorem-proving capabilities of DeepSeek-Math-V2, with the goal of pushing open-source model reasoning to the extreme. It is currently available only through a temporary API (tool calls are not supported) for community evaluation and research; the endpoint is valid until 15:59 UTC on December 15, 2025, with pricing the same as V3.2.

Its report card is quite impressive: it achieved gold-medal-level results in IMO, CMO, ICPC World Finals, and IOI 2025. Official benchmark comparison (excerpt):

Benchmark	GPT-5 High	Gemini-3.0 Pro	V3.2 Thinking	V3.2 Speciale
AIME 2025	94.6	95.0	93.1	96.0
HMMT Feb 2025	88.3	97.5	92.5	99.2
IMOAnswerBench	76.0	83.3	78.3	84.5
LiveCodeBench	84.5	90.7	83.3	88.7
CodeForces	2537	2708	2386	2701
GPQA Diamond	85.7	91.9	82.4	85.7
HLE	26.3	37.7	25.1	30.6

As can be seen, Speciale has already overtaken GPT-5 High on math competition tasks, approaching and in some cases exceeding Gemini-3.0 Pro; however, on knowledge-intensive or "Humanity's Last Exam"-type benchmarks such as GPQA and HLE, it still lags behind Gemini-3.0 Pro. This also confirms its positioning as a "lopsided genius": specializing in highly difficult mathematics, academic logic verification, and competitive programming, rather than being an all-rounder.

III. The Real Highlight: Thinking in Tool-Use

Compared with topping leaderboards, I think the most noteworthy aspect this time is that V3.2 has become DeepSeek's first model to integrate "thinking" directly into tool calls, while also supporting tool calls in both thinking and non-thinking modes.

To this end, the official team built a large-scale Agent training data synthesis method covering 1,800+ environments and 85,000+ complex instructions. This means the model does not merely "know how to think," but can think while acting during multi-step tool calls—this is the prerequisite for Agents to truly be deployed, and the infrastructure for the "digital workforce" (automatically generating reports, writing code, and running processes).

IV. How to Choose?

Daily development / Product integration: Use V3.2 directly; it is compatible with how the original V3.2-Exp was called, so you only need to change the model name, and both cost and length are more favorable.
Mathematics, algorithm competitions, scientific research validation: You can try the Speciale temporary endpoint, but note that it has high token consumption, does not support tool calls, and will be discontinued after December 15.
Building Agent applications: Focus on evaluating the thinking + tool-use combination in V3.2; this is the most practical capability improvement in this update.

V. A Few Observations

From V3 to V3.2, DeepSeek's direction is becoming increasingly clear: benchmark against the strongest closed-source models with open-source models, while treating "cost-effectiveness" and "Agent usability" as core selling points. It does not seek to win across every leaderboard, but instead splits reasoning capability and engineering deployment into two product lines—one responsible for earning a living, the other for looking up at the stars. For developers, this is probably one of the most noteworthy open-source options right now.