DeepSeek-V3.2: Pushing the Frontier of Open Large Language Models
by DeepSeek-AI
How fine-grained sparse attention and scaled RL got an open model into frontier territory.
Overview
DeepSeek-V3.2 is DeepSeek-AI's December 2025 technical report (arXiv:2512.02556, cs.CL, 264 listed authors) on an open-weight model built for long-context efficiency and agentic reasoning. The paper runs Introduction, Architecture, Post-Training, Evaluation and Conclusion, with appendices on MHA and MQA modes, the cold-start template, non-thinking evaluation, competition evaluation methodology and the author list. The architecture section introduces DeepSeek Sparse Attention (DSA), which puts a lightning indexer — a small-head, FP8 scoring module — in front of attention to select roughly 2,048 key-value tokens per query out of a 128K context, then reports a parity evaluation showing output quality holds and an inference-cost analysis showing where the savings land. Continued training in the sparse regime consumed about 943.7B tokens across 15,000 steps. Post-training scales reinforcement learning compute; the released model reports AIME 2025 at 93.1%, HMMT February 2025 at 92.5%, GPQA Diamond at 82.4%, a Codeforces rating of 2386, SWE-Bench Verified at 73.1% and BrowseComp at 51.4%, rising to 67.6% with context management. A high-compute variant, DeepSeek-V3.2-Speciale, is trained on reasoning data with reduced length penalties, drops tool-calling support, and is claimed to reach gold-medal performance at both IMO 2025 and IOI 2025. The third contribution is an agentic task synthesis pipeline generating tool-use training data at scale. The 685B-parameter weights are MIT-licensed on Hugging Face and runnable through vLLM and SGLang, and the DSA kernels are open — indexer logits in DeepGEMM, sparse attention in FlashMLA — so the paper is the reference text for an implementation you can actually run.
At a Glance
- Topic
- Models
- Level
- Advanced
- Format
- Paper
- Cost
- Free
- Duration
- ~2-3 hours for a careful read; longer with the appendices
- Provider
- DeepSeek-AI
- Hands-on
- No
- Certificate
- None
What You’ll Learn
- ✓How a lightning indexer selects top-k tokens for sparse attention
- ✓Why fine-grained sparsity preserves quality where coarser schemes degrade it
- ✓How DSA changes long-context inference cost at 128K context length
- ✓What a scaled reinforcement learning post-training pipeline looks like end to end
- ✓How agentic tool-use training data can be synthesised at large scale
- ✓Which benchmarks the authors chose and how each one was actually scored
- ✓How context management lifts BrowseComp from 51.4 to 67.6 percent
Highlights
- •A detailed public account of fine-grained sparse attention trained at frontier scale
- •Reports the sparse-stage training budget — 943.7B tokens over 15,000 steps — not just results
- •The kernels are open source: DeepGEMM for the indexer, FlashMLA for sparse attention
- •Appendices document competition evaluation methodology, so the IMO and IOI claims are checkable
- •MIT-licensed 685B weights ship alongside the paper and run under vLLM or SGLang
Who It’s For
Best For
- ✓Inference engineers evaluating sparse attention for long-context serving
- ✓Researchers designing reinforcement learning pipelines for reasoning models
- ✓Teams benchmarking open-weight models against closed frontier systems
- ✓Anyone building agents who needs synthetic tool-use data at scale
Prerequisites
- •Working knowledge of transformer attention, KV caching and mixture-of-experts architectures
- •Familiarity with reinforcement learning for LLM post-training such as GRPO or PPO
- •Comfort reading benchmark tables and ablation studies critically
FAQ
What is DeepSeek-V3.2: Pushing the Frontier of Open Large Language Models?
The technical report behind DeepSeek-V3.2, submitted to arXiv on 2 December 2025 by DeepSeek-AI. It documents three contributions: DeepSeek Sparse Attention, a fine-grained mechanism that cuts long-context compute; a scalable reinforcement learning framework whose high-compute variant DeepSeek-V3.2-Speciale reports gold-medal results at IMO 2025 and IOI 2025; and a large-scale pipeline for synthesising agentic tool-use training data. The weights ship under MIT alongside it.
Is DeepSeek-V3.2: Pushing the Frontier of Open Large Language Models free?
DeepSeek-V3.2: Pushing the Frontier of Open Large Language Models is free to access.
What level is DeepSeek-V3.2: Pushing the Frontier of Open Large Language Models for?
DeepSeek-V3.2: Pushing the Frontier of Open Large Language Models is aimed at a advanced audience. Recommended background: Working knowledge of transformer attention, KV caching and mixture-of-experts architectures, Familiarity with reinforcement learning for LLM post-training such as GRPO or PPO, Comfort reading benchmark tables and ablation studies critically.
How long does DeepSeek-V3.2: Pushing the Frontier of Open Large Language Models take?
Expect roughly ~2-3 hours for a careful read; longer with the appendices. Most learners work through it at their own pace.
What will I learn from DeepSeek-V3.2: Pushing the Frontier of Open Large Language Models?
You'll learn: How a lightning indexer selects top-k tokens for sparse attention; Why fine-grained sparsity preserves quality where coarser schemes degrade it; How DSA changes long-context inference cost at 128K context length; What a scaled reinforcement learning post-training pipeline looks like end to end; How agentic tool-use training data can be synthesised at large scale; Which benchmarks the authors chose and how each one was actually scored; How context management lifts BrowseComp from 51.4 to 67.6 percent.
Topics
Sources
This page was written from 3 sources, 1 on domains other than arxiv.org.