DeepSeek-V3 parameter count
DeepSeek-V3 has 671 billion total parameters, with 37 billion activated for each token. DeepSeek states these figures in its official model announcement and repository. V3 is a Mixture-of-Experts (MoE) model, so the active count is the portion used for a token’s computation—not the model’s total parameter count. (deepseek.com)
These figures describe the DeepSeek-V3 model, not every later model in the V3 family. DeepSeek-V3.1, V3.2, and V3.2-Exp are distinct versions; don’t substitute one version’s configuration values for another’s.
Total parameters vs. active parameters
- 671B total parameters describes the full parameter set across the model.
- 37B activated per token describes the parameters DeepSeek says are activated for each token in its MoE architecture. (
github.com)
The active figure helps explain the architecture, but it is not a promise that a deployment needs memory for only 37B parameters. A hosted API or a self-hosted setup has its own serving and infrastructure requirements. Parameter counts alone also don’t establish latency, throughput, or API cost; those depend on the provider and how the model is served.
Practical interpretation: use 671B when you need the total model size; use 37B when discussing the reported per-token activation. Neither number, by itself, tells you whether a particular API will meet your budget or performance target.
Architecture and related specifications
DeepSeek describes V3 as an MoE model using DeepSeekMoE and Multi-head Latent Attention (MLA). The model announcement also reports training on 14.8 trillion tokens. (github.com)
MLA is relevant to memory use during inference: a technical overview explains that it compresses key and value representations for storage in the KV cache, then projects them back for use. That explanation is qualitative; it does not supply a V3-specific memory requirement or a guaranteed serving-performance figure. (github.com)
Keep architecture figures separate from API generation settings. A parameter count describes the model; controls such as output limits or sampling settings describe how a particular API request is configured.
Production API availability: check the served model, not just the weights
As of November 10, 2026, DeepSeek’s current API quick-start and pricing pages list deepseek-flash and deepseek-v4-pro, not DeepSeek-V3. The quick-start says deepseek-flash is served by DeepSeek-V4.1-Flash-1017. The current pricing page likewise identifies the served versions as DeepSeek-V4.1-Flash-1017 and DeepSeek-V4-Pro-0813. So the published availability of V3 weights does not mean V3 is currently offered as a model choice in DeepSeek’s direct API. Check the provider’s live model list before building against a model ID, since API catalogs can change. (api-docs.deepseek.com)
DeepSeek documents OpenAI-compatible access at https://api.deepseek.com, Bearer-token authentication, and a current model list with metadata such as context window, output limit, and supported capabilities. Those details apply to the models listed for that API; they do not establish V3 availability. (api-docs.deepseek.com)
Token Forge Cloud describes its managed APIs as including DeepSeek models, but the available product information here does not confirm a current V3-specific API identifier or its limits. For a production decision, verify the exact served version, request controls, context and output limits, and commercial terms with the provider before integrating. (tokenforgecloud.com)
Bottom line
DeepSeek-V3: 671B total parameters; 37B activated per token. Those are the official model figures. For production API selection, distinguish the open V3 model from the model versions a provider currently serves: DeepSeek’s direct API documentation currently lists V4-family models, while a V3-specific Token Forge Cloud offering is not confirmed by the product evidence cited here.
Sources
DeepSeek-V3 — DeepSeek AI GitHub repository
Introducing DeepSeek-V3 — DeepSeek
A Technical Tour of the DeepSeek Models from V3 to V3.2
Your First API Call — DeepSeek API Docs
Models & Pricing — DeepSeek API Docs
Lists Models — DeepSeek API Docs