Direct answer: how big is DeepSeek-V3.2?
If “deepseek-v3 2” means DeepSeek-V3.2, the Hugging Face listing labels it as a 685-billion-parameter model. A pinned repository snapshot lists 690 GB and shows 163 Safetensors weight shards. The repository is tagged FP8; the 690 GB figure applies to that listed snapshot, not to every format or revision. Parameter count and downloadable repository size are different measures. Neither number, by itself, specifies the RAM or GPU memory needed to run the model.
| Meaning of “size” | Figure or information available | What it does not tell you |
|---|---|---|
| Parameter count | 685B, as labeled on the model listing | A published active-parameter count |
| Download footprint | 690 GB for the pinned FP8 Safetensors repository snapshot; 163 weight shards | A universal file size across formats or revisions, or a minimum RAM/VRAM requirement |
| Runtime memory | No verified minimum or conditioned deployment estimate established here | Whether a particular hardware setup meets a context-length or speed target |
DeepSeek describes V3.2 as the official successor to V3.2-Exp. They are distinct model versions, so don’t carry file or hardware figures from one over to the other.
Local hosting considerations
Treat 690 GB as the listed repository download footprint for that pinned FP8 snapshot, not as a RAM or VRAM specification. A single 24 GB GPU cannot contain the listed weight files in full at that size. That comparison does not establish whether a particular quantized, multi-GPU, or CPU-offloaded deployment will work.
The model repository includes example serving instructions for vLLM. Those instructions show how to launch a server, but the available excerpt does not establish a hardware minimum or a validated configuration with specified memory, context length, and performance.
A community thread about V3.2-Exp, not V3.2, contains conflicting user estimates: 750 GB RAM, and 400 GB RAM when quantized plus about 96 GB VRAM for “decently fast” operation. The commenters do not provide enough comparable setup details to treat those numbers as verified requirements.
So the useful local answer is limited: plan for substantial storage for the listed snapshot, but don’t infer a RAM or VRAM requirement from repository size alone. The exact hardware feasibility, including the effects of quantization, offload, serving software, context length, and speed target, remains unestablished by the cited material.
Local deployment or a production API?
With local hosting, you download and operate the model weights and serving stack yourself. With a hosted API, the provider operates the model-serving infrastructure, so you do not need to download and host the weights. The trade-off is that you must verify the provider’s exact model ID and version, capabilities, and operational terms for your workload.
DeepSeek’s documented quick-start names deepseek-flash and deepseek-v4-pro. That quick-start alone does not establish whether V3.2 is available through an API or how those model IDs map to versions. Check the current model catalog and confirm the exact model ID and version before relying on a hosted option. Token Forge Cloud describes production APIs for DeepSeek models generally, but the available product information does not confirm a V3.2-specific model ID. If considering Token Forge or another provider, verify the exact model and version directly before relying on it.
Bottom line
For the likely intended model, the quick answer is 685B parameters and 690 GB for a pinned FP8 repository snapshot. The local serving instructions do not establish a hardware minimum, and the V3.2-Exp memory anecdotes are not V3.2 requirements. A hosted API avoids operating the weights yourself, but confirm the provider’s current model ID and version before choosing it.
Sources
DeepSeek-V3.2 pinned repository snapshot · Hugging Face
DeepSeek-V3.2 Release · DeepSeek API Docs
How Much VRAM? · DeepSeek-V3.2-Exp discussion
Your First API Call · DeepSeek API Docs