deepseek
DeepSeek: DeepSeek V3.2 Exp
deepseek/deepseek-v3.2-exp
Our take
AI-assisted analysis based on the sources below. Suitability is an editorial assessment; test it on your own workload.
DeepSeek-V3.2-Exp is an experimental text-to-text model that DeepSeek describes as an intermediate step between V3.1-Terminus and its next-generation architecture. Its main change is DeepSeek Sparse Attention, a mechanism intended to improve long-context training and inference efficiency while keeping output quality close to V3.1-Terminus, with the developer reporting broad benchmark parity. The OpenRouter catalog lists a 163,840-token context window plus tool calling, structured outputs, and optional reasoning controls. Weights are released under the MIT license, with documented vLLM and SGLang support.[1][2]
Strengths
- DeepSeek Sparse Attention is designed to make long-context training and inference substantially more efficient while preserving near-identical output quality, according to the developer.[2]
- The developer reports performance on par with V3.1-Terminus across reasoning, coding, and agentic tool-use benchmarks.[2]
- The OpenRouter catalog lists a 163,840-token context window, tool calling, structured outputs, and configurable reasoning parameters.[1]
Limitations
- It is an explicitly experimental release meant to validate an architectural research direction, so its long-term support and positioning are uncertain.[2]
- An earlier inference demo contained a RoPE implementation discrepancy that could degrade performance; it is fixed, but self-hosters should use the updated code.[2]
- Evidence is limited to the developer's model card and one aggregator catalog with no independent evaluations, so real-world quality should be validated on your own workloads.[1][2]
Where to use it
- Long-context applications such as extended document analysis or multi-turn conversations, where the sparse attention design may reduce compute costs; worth piloting given the efficiency claims.[1][2]
- Agentic workflows that rely on tool calling and structured outputs, which the catalog lists as supported parameters.[1]
- Self-hosted deployments for teams wanting MIT-licensed weights, using the documented vLLM, SGLang, or Hugging Face inference paths.[2]
Sources
- [1] OpenRouter catalog · DeepSeek: DeepSeek V3.2 ExpSources checked: Nov 13, 2026
- [2] Model documentation · DeepSeek: DeepSeek V3.2 ExpSources checked: Nov 13, 2026
About this model
Description below is the original text supplied by OpenRouter, shown in the source language.
DeepSeek-V3.2-Exp is an experimental large language model released by DeepSeek as an intermediate step between V3.1 and future architectures. It introduces DeepSeek Sparse Attention (DSA), a fine-grained sparse attention mechanism...
Specifications
- Model ID
- deepseek/deepseek-v3.2-exp
- Creator
- deepseek
- Added to OpenRouter
- 29.09.2025
- Context length
- 163,840
- Modality
- text->text
- Input modalities
- text
- Output modalities
- text
- Token pricing
- USD per 1M tokens: in $0.270 / out $0.410
- Knowledge cutoff
- 31.07.2025
- Expiration date
- —
Tasks and capabilities
A grade appears only where the source supplies a matching measurement for that task, the model is active and the cohort holds at least 10 comparable models. The capability column lists facts the source declares — supported formats, reasoning and tool parameters — which are not measured quality.
| Task | Grade | Source-declared capability |
|---|---|---|
| General quality | No measurement | reasoning: optional · parameters: include_reasoning, reasoning |
| Coding | No measurement | — |
| Agents | No measurement | parameters: structured_outputs, tool_choice, tools |
| Writing | No measurement | output: text |
| Reasoning | No measurement | reasoning: optional · parameters: include_reasoning, reasoning |
| Document analysis | No measurement | — |
| Vision | No measurement | — |
| Image generation | No measurement | — |
| Audio | No measurement | — |
| Video | No measurement | — |
| Embeddings | No measurement | — |
Evaluations
| Source | Benchmark | Value | Grade | Data retrieved |
|---|---|---|---|---|
| Design Arena | models · 3dwin rate 56.4% · rank 77 | 1,131Elo | BB — next 30% of the cohort | 13.11.2026 |
| Design Arena | models · codecategorieswin rate 54.2% · rank 78 | 1,161Elo | BB — next 30% of the cohort | 13.11.2026 |
| Design Arena | models · datavizwin rate 50.6% · rank 83 | 1,150Elo | CC — below the B band | 13.11.2026 |
| Design Arena | models · gamedevwin rate 53% · rank 80 | 1,135Elo | CC — below the B band | 13.11.2026 |
| Design Arena | models · svgwin rate 42% · rank 87 | 1,023Elo | CC — below the B band | 13.11.2026 |
| Design Arena | models · uicomponentwin rate 53.3% · rank 77 | 1,159Elo | CC — below the B band | 13.11.2026 |
| Design Arena | models · websitewin rate 54.3% · rank 81 | 1,172Elo | BB — next 30% of the cohort | 13.11.2026 |
How grades work
Each letter grade compares a model only against other models scored by the same source and metric in the current snapshot. S means the top 10%, A the next 20%, B the next 30%, C the rest. Models without a published measurement are shown as not rated — a supported format alone never implies measured quality.
- S — S — top 10% of the cohort
- A — A — next 20% of the cohort
- B — B — next 30% of the cohort
- C — C — below the B band
Providers
Optional live provider statistics from OpenRouter. Missing values are unknown, not zero.
| Provider | Context length | Price in /1M | Price out /1M | Latency p50 (30m, s) | Throughput p50 (30m, tok/s) | Uptime (30m) |
|---|---|---|---|---|---|---|
| SiliconFlowfp8 | 163,840 | $0.270 | $0.410 | — | — | 100% |
Provider data retrieved: 13.11.2026
Related articles
No published articles identify this model yet.