Which DeepSeek V3 should you run with Ollama?
For the full DeepSeek V3 model, Ollama’s library lists deepseek-v3. The default listing shows a 671-billion-parameter model, a 404GB download, and a 160K context window. Its README says Ollama 0.5.5 or later is required. These details correspond to the listed default 671B Q4_K_M variant; they are not a guarantee that it will fit or run well on your computer. (Ollama model page, available tags)
A smaller listing with “DeepSeek V3” in its name is not the same model: xiaowangge/deepseek-v3-qwen2.5:7b-fp16 describes itself as based on DistilQwen2.5-DS3-0324-Series. It lists 7.62B parameters, a 15GB artifact, and an Ollama 0.6.8-or-later requirement. Treat it as a distilled Qwen2.5-related option—not the full DeepSeek V3 checkpoint. (model listing)
| If you need… | Consider… | Important distinction |
|---|---|---|
| The full DeepSeek V3 model in Ollama | deepseek-v3 |
The default listing shows 671B parameters, a 404GB download, and a 160K context window. Check storage and hardware before starting. (Ollama tags) |
| A smaller, V3-related model to try locally | xiaowangge/deepseek-v3-qwen2.5:7b-fp16 |
The listing describes it as based on a distilled Qwen2.5 series; it is not the full V3 model. (model listing) |
| DeepSeek V3 through DeepSeek’s hosted API | V3 availability is not established by the API documentation reviewed here | The current API documentation identifies deepseek-flash as DeepSeek-V4.1-Flash-1017. That documents a different model; it does not establish whether another V3 API option exists. (API quick start, models and pricing) |
Run the full model locally
Make sure Ollama is installed and that you have enough disk space for the listed download. The default deepseek-v3 listing specifies Ollama 0.5.5 or later and shows a 160K context window for the 404GB Q4_K_M artifact. It does not specify minimum RAM or VRAM, a tested machine configuration, or expected speed. Artifact size is not a reliable estimate of the memory needed while running the model. (Ollama model page, available tags)
Run the model:
ollama run deepseek-v3
To send a chat request to Ollama’s local API, use the model name in the request body:
curl http://localhost:11434/api/chat \
-d '{
"model": "deepseek-v3",
"messages": [{"role": "user", "content": "Hello!"}]
}'
The Ollama model page provides this localhost endpoint and request format. This example assumes Ollama is running on the same computer and the model is available locally; localhost:11434 is not a hosted production endpoint. (Ollama model page)
If the full model is too large for your use case, the smaller distilled listing has its own run command:
ollama run xiaowangge/deepseek-v3-qwen2.5:7b-fp16
That listing says Ollama 0.6.8 or later is required. Its 15GB artifact can help with download and storage planning, but does not establish minimum runtime memory or performance. (model listing)
Ollama locally or a production API?
Choose local Ollama when you want one of these local model variants, can accommodate its storage and runtime requirements, and are prepared to operate the inference service yourself. Before committing to the full V3 model, confirm that your hardware can load it; the Ollama listing does not give a tested minimum configuration.
Choose a hosted API only if its documented model matches your requirement. DeepSeek’s API documentation currently identifies deepseek-flash as served by DeepSeek-V4.1-Flash-1017. That is not DeepSeek V3. The documentation reviewed here does not establish whether a V3 API option is available through another identifier or route, so confirm with the provider before substituting a different model. (API quick start, models and pricing)
For any hosted API you are considering, verify the model identifier, endpoint, authentication, required features, limits, data terms, and price for your region in that provider’s documentation. TokenForge V3 support and service terms are not established here, so this article cannot recommend it as a V3 API provider.
What not to confuse with current support
A Hugging Face discussion from December 2024 and January 2025 includes a failed attempt to pull deepseek-ai/DeepSeek-V3 and disagreement over whether some reported smaller DeepSeek models were actually V3 or DeepSeek 2.5. That is historical community discussion, not current evidence of Ollama availability. Use the current Ollama listing—not the old discussion—to identify the library command. (Hugging Face discussion, Ollama model page)
Bottom line: For the full model in Ollama, the listed command is ollama run deepseek-v3; the default 671B Q4_K_M listing shows a 404GB download and a 160K context window. Verify your machine can handle it. The 7B-labelled Qwen2.5-derived listing is a distinct, smaller V3-related option. DeepSeek’s reviewed API documentation identifies deepseek-flash as V4.1-Flash-1017, not V3; V3 API availability is not established by those documents.
Sources
xiaowangge/deepseek-v3-qwen2.5:7b-fp16 — Ollama
Your First API Call — DeepSeek API Docs
Models & Pricing — DeepSeek API Docs
Has anyone tried running this model on Ollama? — Hugging Face