What is DeepSeek-V3.2?
DeepSeek-V3.2 is a 685.4B parameter open-weights model from DeepSeek-AI, published on Hugging Face on December 1, 2025 under the MIT license. DeepSeek builds it on three things: DeepSeek Sparse Attention (DSA), an attention mechanism that cuts computational complexity for long-context work, a scaled reinforcement learning framework for post-training, and a synthesis pipeline that generates agentic training data at scale. The weights are free to download and MIT licensed, but this is a server-class model: the smallest quantized build is 161 GB, so plan for a multi-GPU box or a very large unified-memory machine.
| Specification | DeepSeek-V3.2 |
|---|---|
| Total parameters | 685.4B |
| Architecture | Same model structure as DeepSeek-V3.2-Exp, with DeepSeek Sparse Attention |
| Original precision | FP8 (F8_E4M3) weights |
| Modalities | Text |
| Reasoning | Thinking mode, including thinking with tools |
| Recommended sampling | temperature 1.0, top_p 0.95 |
| Release date | December 1, 2025 |
| License | MIT |
The chat template changed a lot in this release. Tool calling has a revised format, the model can think while it calls tools, and a new developer role exists solely for search agent scenarios. There is no Jinja template in the repo; DeepSeek ships Python encoding scripts and test cases instead, so check your serving stack handles the new format before you rely on tool calls.
What DeepSeek-V3.2 is good at
DeepSeek publishes no benchmark numbers in text on this model card, only a benchmark image, so what follows are the vendor's own claims. DeepSeek reports that scaling reinforcement learning post-training brings V3.2 to performance comparable to GPT-5, and that the high-compute variant, DeepSeek-V3.2-Speciale, surpasses GPT-5 with reasoning on par with Gemini-3.0-Pro. The concrete evidence offered: gold-medal performance at the 2025 International Mathematical Olympiad and the International Olympiad in Informatics, with the model's selected submissions for IMO 2025, IOI 2025, the ICPC World Finals and CMO 2025 published in the repo for the community to verify.
The other focus is agents. The synthesis pipeline exists to push reasoning into tool-use scenarios, and DeepSeek says it improves compliance and generalization in complex interactive environments. Two practical notes from the card: the Speciale variant is built exclusively for deep reasoning and does not support tool calling, and for local deployment DeepSeek recommends temperature 1.0 with top_p 0.95.
DeepSeek-V3.2 hardware requirements
The system requirement to check is memory. DeepSeek ships the original FP8 weights; the GGUF builds below come from unsloth/DeepSeek-V3.2-GGUF. UD-TQ1_0 ships as one single file; the other four are split into shards, so the size below is the total across a build's files:
| Memory | Build to pick | File size |
|---|---|---|
| 192 GB | UD-TQ1_0 | 161.28 GB |
| 256 GB | UD-IQ2_M | 228.11 GB |
| 320 GB | UD-IQ3_XXS | 273.12 GB |
| 384 GB | IQ4_XS | 358.37 GB |
| 512 GB and up | UD-Q4_K_XL | 407.81 GB |
When two builds both fit, take the larger one, and leave headroom above the file size for the KV cache and the rest of the system. If the GGUF format is new to you, start with what GGUF is.
How to run DeepSeek-V3.2 in Atomic Chat
Atomic Chat is a free local app for macOS, Windows and Linux. It includes a Hugging Face model browser and a built-in chat, with no manual llama.cpp build required.
- Download Atomic Chat for your platform and open it.
- Search for DeepSeek-V3.2 in the model browser and open Download Options.
- Pick the build that fits the memory you have, then start a chat.
If your machine is smaller than the table above, see every DeepSeek model you can run locally, including DeepSeek-R1.
DeepSeek-V3.2 license
DeepSeek-V3.2 is released under the MIT license, and the card states that it covers both the repository and the model weights. That permits commercial use, modification and redistribution with no royalties, so you can build products on the model and serve it on your own hardware without a usage fee.