DeepSeek-V3.2
DeepSeek's MIT-licensed 685B reasoning model with sparse attention for cheaper long context; the V4 family has since replaced it.
In short. DeepSeek-V3.2 is a very large reasoning model from DeepSeek that anyone may download under the MIT license. It was built to be cheaper to run on long inputs and to think while using tools. DeepSeek's newer V4 models have replaced it, but it is still one of the most downloaded large open models.
Summary
DeepSeek released V3.2 on 1 December 2025 as the official successor to the V3.2-Exp experiment from September. The checkpoint is a mixture-of-experts model with about 685 billion parameters: 61 layers and 256 routed experts, with 8 active per token plus one shared expert. That is the same layout as DeepSeek-V3, which has about 37 billion active parameters. The main change is DeepSeek Sparse Attention, which lets each token look at a selected part of the context and lowers the cost of long inputs. The weights are published in FP8, about 690 GB, and the configuration allows 163,840 positions.
DeepSeek's report names three ideas. The first is sparse attention. The second is a scaled-up reinforcement-learning stage, which DeepSeek says brings V3.2 to GPT-5-level results in its tests. The third is a pipeline that synthesizes agent training tasks, covering more than 1,800 environments and 85,000 complex instructions, so that the model can think while it calls tools. A high-compute variant, DeepSeek-V3.2-Speciale, is also published. DeepSeek says it rivals Gemini 3.0 Pro on reasoning and reached gold-medal level at the 2025 IMO, CMO, ICPC World Finals and IOI, at the price of using many more tokens. It was API-only and had no tool calling at launch. All of this is self-reported, and DeepSeek released its competition submissions for others to check.
There is no Jinja chat template. Use the Python scripts in the repository's encoding folder, whose output parser handles only well-formed text and is not meant for production as is. The license is MIT. DeepSeek moved on with the V4 preview in April 2026 and later V4 and V4.1 releases, so compare those first for new work.
Good for
- Self-hosting a widely supported large reasoning model under the MIT license
- Agent workflows that need the model to think between tool calls
- Studying DeepSeek Sparse Attention and its long-context cost savings
- Checking the published competition submissions
Watch out for
- DeepSeek V4 and V4.1 are newer, so compare them before starting new deployments
- No Jinja chat template is supplied; you need DeepSeek's encoding scripts
- It needs a multi-GPU node, with about 690 GB of FP8 weights
- Speciale uses far more tokens and had no tool calling at launch
- It is text-only, and the benchmark claims come from DeepSeek
Performance claims come from the developer and have not been independently verified.
How much memory does it need?
Estimates for the V3.2 variant. A mixture-of-experts model must hold all of its experts in memory even though only some are used for each token, so total parameters decide the memory and active parameters decide the speed.
| Variants | Parameters | BF16 | FP8 / INT8 | 4-bit (Q4) | Fits on |
|---|---|---|---|---|---|
| V3.2 FP8 | 685B 37B active | 1.58 TB | 788 GB | 473 GB | Big-memory desktop |
| V3.2-Speciale FP8 | 685B 37B active | 1.58 TB | 788 GB | 473 GB | Big-memory desktop |
Estimates: weights plus about 15% overhead. Real usage grows with context length and depends on the software.
The license in plain words
The shortest and simplest permissive license. You can do almost anything with the model as long as you keep the copyright and license notice.
Frequently asked questions
- What is DeepSeek-V3.2?
- DeepSeek-V3.2 is a very large reasoning model from DeepSeek that anyone may download under the MIT license. It was built to be cheaper to run on long inputs and to think while using tools. DeepSeek's newer V4 models have replaced it, but it is still one of the most downloaded large open models.
- Can I use DeepSeek-V3.2 commercially?
- Yes. DeepSeek-V3.2 is released under MIT, a permissive license. You can use it commercially if you follow the standard conditions, such as keeping the license notice.
- Is DeepSeek-V3.2 open source?
- DeepSeek-V3.2 is released under MIT, a license approved by the Open Source Initiative. Training data and training code may not be public, so it is best described as open weights under an open-source license.
- Can I run DeepSeek-V3.2 on my own computer?
- Only on a machine with a lot of memory. The smallest variant, V3.2, needs about 473 GB at 4-bit precision. Desktops with 128 to 512 GB of unified memory can hold it, at modest speed.
- When was DeepSeek-V3.2 released?
- DeepSeek-V3.2 was released by DeepSeek on December 1, 2025.
- How long a context window does DeepSeek-V3.2 have?
- DeepSeek-V3.2 supports a context window of up to 160K tokens.
Official links
Sources
Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.
- DeepSeek-V3.2 Release (DeepSeek API docs, 1 Dec 2025) api-docs.deepseek.com
- DeepSeek-V3.2 model card (Hugging Face) huggingface.co
- DeepSeek-V3.2-Speciale model card (Hugging Face) huggingface.co
- DeepSeek-V3.2: Pushing the Frontier of Open Large Language Models (technical report) arxiv.org
Release date and facts last verified November 11, 2026.
Related models
| Model | Released | Parameters | Context | License | Smallest variant needs |
|---|---|---|---|---|---|
| DeepSeek V4 DeepSeek | 284B – 1.6T 49B active | 1M | MIT (Permissive) | 176 GB Big-memory desktop | |
| DeepSeek V4.1 Flash DeepSeek | 552B 16B active | 1M | MIT (Permissive) | 561 GB GPU server | |
| GLM-5.2 Z.ai (Zhipu AI) | 753B 40B active | 1M | MIT (Permissive) | 520 GB GPU server | |
| Kimi K3 Moonshot AI | 2.8T 104B active | 1M | Kimi K3 License (Conditional) | 1.72 TB Multi-node cluster |