Short answer
For DeepSeek-V3.2, Modular lists 685B total parameters and 37B active parameters. The active figure is not a second estimate of the model’s total size: it describes the active parameters Modular reports for each token.
Keep the experimental variant separate. Together AI lists 685B for DeepSeek-V3.2-Exp, but its listing does not say whether that figure means total or active parameters. Don’t transfer that unspecified figure—or specifications for one release—to another variant.
| Model | Reported parameter information | What the source establishes |
|---|---|---|
| DeepSeek-V3.2 | 685B total; 37B active | Modular explicitly identifies 685B as total and lists 37B active. |
| DeepSeek-V3.2-Exp | 685B; total-versus-active type unspecified | Together AI displays 685B for this exact variant, without defining the count type. |
| Original DeepSeek-V3 | Not compared here | This is a distinct model name; don’t assume V3.2 figures apply to it. |
What the architecture details tell you
The DeepSeek-V3.2 configuration lists 256 routed experts, one shared expert, and eight experts per token. Those are architecture settings, not a substitute for a stated total or active parameter count; they should not be used alone to calculate either figure.
Parameter counts also do not, by themselves, tell you a hosted API’s price, speed, memory use, or serving requirements. Check those with the provider for the specific deployment you plan to use.
Check the exact version before choosing an API
Modular documents a hosted API example for DeepSeek-V3.2. The example uses the model identifier deepseek-ai/DeepSeek-V3.2 and base URL https://model.api.modular.com. Confirm current availability and limits with the provider before relying on that path in production.
DeepSeek’s direct API documentation lists different currently selectable model names. Its quick-start names deepseek-flash and deepseek-v4-pro; it says Flash is served by DeepSeek-V4.1-Flash-1017. The V3.2 release announcement said V3.2 was available through the API at launch, but that announcement does not establish current availability. Verify the current model list before selecting DeepSeek’s direct API for V3.2.
For DeepSeek-V3.2-Exp, Together AI’s listing says the model is not available on Together’s Serverless API. That restriction is specific to Together Serverless; the listing does not establish availability through a dedicated deployment or another provider.
Bottom line: Use 685B total and 37B active when you mean DeepSeek-V3.2, as listed by Modular. For V3.2-Exp, Together AI’s displayed 685B has no total-versus-active definition, and the listing says it is unavailable on Together’s Serverless API. The listing does not establish other deployment options. For production, check the exact variant, request identifier, serving channel, and current provider documentation.
Sources
DeepSeek V3.2: Optimized 685B MoE Inference on Modular
DeepSeek-V3.2-Exp API: Pricing, Benchmarks & Docs | Together AI
config.json · deepseek-ai/DeepSeek-V3.2 at main