benchmarks
DeepSeek-V3.2 is an open-weight model supporting reasoning, coding and tool-assisted workflows, with a documented 128K-token context window. Its technical report describes stronger agent capabilities and more efficient long-context attention, while identifying knowledge breadth and reasoning-token efficiency as remaining limitations.
Benchmark results
Every figure links to where it was published. Compare like with like — harnesses and prompt setups differ between vendors.
| Benchmark | Score | Source |
|---|---|---|
| AIMECompetition mathematics | 93.1%Dec 2025 | View |
| GPQA DiamondPhD-level science questions | 82.4%Dec 2025 | View |
| Humanity's Last Exam — without tools | 25.1%Dec 2025 | View |
| LiveCodeBenchFresh competitive-programming problems | 83.3%Dec 2025 | View |
| SWE-bench VerifiedResolving real GitHub issues | 73.1%Dec 2025 | View |
| Terminal-BenchLong-horizon command-line agent tasks | 46.4%Dec 2025 | View |
| LMArena TextHuman preference Elo from blind chat votes (not a percentage) | 1425 Elo22 Oct 2026 | View |
| Aider PolyglotCode editing across six programming languages | 74.2% | View |
| LMArena WebDevHuman preference Elo for building web apps (not a percentage) | 1362 Elo23 Oct 2026 | View |
Weighted mean of 7 verified benchmark results: AIME 93.1%, GPQA Diamond 82.4%, Humanity's Last Exam — without tools 25.1%, LiveCodeBench 83.3%, SWE-bench Verified 73.1%, Terminal-Bench 46.4%, Aider Polyglot 74.2%.
Capabilities
- Supports thinking and non-thinking modes.
- Integrates reasoning with tool calls in multi-step workflows.
- Demonstrates mathematical and scientific reasoning in published vendor evaluations.
- Supports code generation, repository-level software repair and terminal-based tasks.
- Provides open weights for deployment on suitably provisioned infrastructure.
What improved
- Extends thinking into tool-use workflows, improving continuity between reasoning and external actions.
- Uses large-scale synthetic agent-task training to strengthen instruction following and tool-use generalization.
- Scales reinforcement learning to improve reasoning and agent performance.
- Builds on V3.2-Exp's DeepSeek Sparse Attention, reducing long-context attention computation relative to V3.1.
Known flaws & limitations
- The technical report identifies breadth of world knowledge as a remaining gap relative to leading proprietary models.
- Difficult tasks can require lengthy reasoning outputs, limiting token efficiency.
- Full-model self-hosting requires substantial infrastructure; sparse activation does not remove the need to store model weights.
- Published agent results depend on evaluation scaffolding, tools and execution settings, not model weights alone.
Risks, safety & security issues
- DeepSeek's hosted infrastructure experienced a publicly documented database exposure in January 2025; this predates V3.2 and is not evidence of a V3.2 model vulnerability.
- Tool-enabled deployments require external permission checks and sandboxing because model-generated actions can affect files, credentials and connected systems.
- Knowledge limitations and imperfect benchmark performance make unreviewed outputs unsuitable as authoritative evidence in high-stakes decisions.
- Sensitive hosted deployments should establish contractual requirements for data handling, retention and residency before submitting confidential information.
Good fit for
- Mathematical reasoning and technical problem-solving with answer verification.
- Coding assistance and repository maintenance with tests and human review.
- Tool-assisted research and software workflows within restricted execution environments.
- Open-weight deployment by teams with sufficient infrastructure and operational expertise.
Poor fit for
- High-stakes factual decisions without independent verification.
- Unsupervised agents with unrestricted access to production systems or sensitive credentials.
- Low-resource full-model self-hosting.
- Strictly latency-sensitive tasks where lengthy reasoning is unacceptable.
References
Profile last reviewed 10/27/2026 and re-checked automatically every few hours against the vendor's published documentation, papers and public leaderboards. Scores come from the cited benchmark results above — confirm limits and pricing against the provider's own documentation before building on them.