DeepSeek-V3.2 Series Open Source, Performance Rivals Gemini-3.0-Pro

🚀 Surprise Release on ChatGPT’s Third Anniversary

On the third anniversary of ChatGPT, DeepSeek suddenly unveiled two new AI models:

• DeepSeek‑V3.2 — designed for balanced practicality, ideal for daily Q&A, general Agent tasks, and real‑world tool invocation.
• 🧠 Reasoning ability: on par with GPT‑5, slightly below Gemini‑3.0‑Pro.
• DeepSeek‑V3.2‑Speciale — built for extreme reasoning, matching Gemini‑3.0‑Pro in benchmarks, excelling in advanced mathematical and logical tasks.
• 📚 Won gold medals in IMO 2025, CMO 2025, ICPC World Finals 2025, and IOI 2025, with ICPC second among humans and IOI tenth among humans.

---

DeepSeek‑V3.2 — Balanced, High‑Performance Agent Model

Key Characteristics

• GPT‑5‑level reasoning
• Shorter outputs than Kimi‑K2‑Thinking → faster response times
• Unified “thinking into tool invocation” — works in thinking and non‑thinking tool modes
• Trained with large‑scale Agent datasets: 1,800+ environments, 85,000+ complex instructions

Benchmark Highlights

• Outperforms existing open‑source models in Agent evaluations.
• Wins tool‑invocation tests without specific training for those tools.

---

DeepSeek‑V3.2‑Speciale — Extreme Reasoning Edition

• Integrates DeepSeek‑Math‑V2 theorem‑proving
• Best for mathematical proofs, programming competitions, academic research
• Not optimized for casual conversation or writing
• Research‑only, no tool invocation
• Outperforms V3.2 in complex tasks → much higher token usage & cost

---

Availability

• App & Web use DeepSeek‑V3.2 by default
• Speciale available via temporary API
• Technical report, paper, and benchmarks released publicly

---

🔍 Technical Innovations in V3.2

DeepSeek Sparse Attention (DSA)

• New attention mechanism
• Traditional: O(L²) complexity → slow for long contexts
• DSA: O(L·k), with k ≪ L
• Faster reasoning for long sequences with no performance drop
• FP8 precision, works with MLA (Multi‑Query Attention)
• Two key components:
• Lightning Indexer — quickly ranks token relevance
• Fine‑grained token selection

---

Training Process — V3.1 Terminus to V3.2

Phase 1: Dense Warm‑up

• Dense attention maintained
• Train lightning indexer to align with main attention
• 1,000 steps, 2.1 B tokens

Phase 2: Introduce Sparse Mechanism

• Each query token selects 2,048 key‑value pairs
• 15,000 steps, 943.7 B tokens

---

🏎 Performance Gains

On 128 k sequences:

| Stage | V3.1‑Terminus Cost | V3.2 Cost |

|-------------|--------------------|-----------|

| Prefill | $0.70 / 1M tokens | ~$0.20 |

| Decode | $2.40 / 1M tokens | $0.80 |

---

💡 Heavy Investment in RL

• RL compute budget >10% of pre‑training — rare in open‑source LLMs
• Developed stable, scalable RL protocol
• Targeted post‑training improvements for complex tasks

---

RL Scalability Enhancements

• Unbiased KL Estimation — removed systemic bias, stabilized gradients
• Offline Sequence Masking — prevents off‑policy degradation
• Keep Routing for MoE — keeps training/inference expert paths consistent

---

Expert Distillation Approach

Task Domains:

• Mathematics
• Programming
• Logical reasoning
• Agent general tasks
• Agent programming
• Agent search

Each domain: supports “thinking” & “non‑thinking” modes

→ Expert models feed high‑quality domain data into final training

---

🔧 Breakthrough in Agent Capabilities

• Combines reasoning + tool usage
• Optimized thinking context management:
• Discards reasoning traces only on new user messages, not tool actions
• Tool‑call history & results preserved

---

Prompt Engineering for Cold Starts

• Special prompts encourage tool calls during reasoning
• For competitions: clearly separates “thinking” from final answer

---

Real‑World Applicability

Platforms like AiToEarn官网 enhance model utility by:

• Multi‑platform publishing (Douyin, Bilibili, LinkedIn, X, etc.)
• AI content creation, analytics, and monetization
• Public AI model rankings (AI模型排名)

---

⚙ Automated Environment Synthesis

• 1,827 environments, 85,000 complex prompts
• Example: travel plan under constraints (budget balance, no repeat cities)
• “Hard to solve, easy to verify” — perfect for RL

---

Code & Search Agent Data

• Code Agent: millions of GitHub issue–PR pairs → executable environments in Python, Java, JavaScript
• Search Agent: multi‑agent pipeline, long‑tail entity sampling, verified answers

Results:

• SWE‑Verified: 73.1%
• Terminal Bench 2.0: 46.4%
• Tool benchmarks: near closed‑source model performance

---

📉 Limitations

• Lower training FLOPs → narrower world knowledge than top closed‑source models
• Less token‑efficient — longer trajectories needed for equal quality
• Improvement targets set for future releases

---

DeepSeek, when will R2 arrive?! 🎯

---

For researchers, developers, and AI creators, DeepSeek‑V3.2 is a blueprint for integrating reasoning, tool usage, and efficiency.

Platforms like AiToEarn官网 let builders:

• Connect AI generation with publishing
• Analyze performance across platforms
• Monetize creativity at scale

Open‑source + multi‑platform reach = real‑world AI impact.