DeepSeek V3
The open-weight 671B-param MoE that put DeepSeek on the frontier map.
DeepSeek has replaced this model with DeepSeek V4.
Context window
128K tokens
Input / 1M tokens
Free
Output / 1M tokens
Free
Provider
DeepSeek
Open-weight (MIT license) - free to self-host. DeepSeek's own API no longer serves V3, and the deepseek-chat and deepseek-reasoner names were fully retired on July 24, 2026 (they last pointed to V4 Flash). A few third-party hosts still serve V3: OpenRouter lists deepseek/deepseek-chat from about $0.26/1M input and $0.71/1M output via DeepInfra. · Data verified 2026-11-08
DeepSeek V3 is an open-weight mixture-of-experts model (671B total / 37B active) first released December 26, 2024 under the MIT license, pre-trained on nearly 15 trillion tokens. It offers a 128K-token context window and strong instruction-following and coding performance, competing with both open and closed models of its era. DeepSeek V4 replaced it as DeepSeek's flagship, and DeepSeek's API no longer serves V3 (the deepseek-chat name was retired on July 24, 2026). The open weights are still downloadable and some third-party hosts still serve it, so choose it only for pinned, reproducible deployments or existing fine-tunes.
Capability index
Relative estimates (0-100) to place this model against its peers, grounded in published benchmarks.
How to access it
Download open weights from Hugging Face (deepseek-ai/DeepSeek-V3) to self-host, or use a third-party host that still serves it, such as OpenRouter (deepseek/deepseek-chat, served by DeepInfra and StreamLake in November 2026). DeepSeek's own API no longer offers V3.
Strengths
- ✓Open weights under MIT license - fully self-hostable and fine-tunable
- ✓Strong general-purpose reasoning, coding, and instruction following for its generation
- ✓128K context window suitable for long documents
- ✓Very low cost via third-party API providers
- ✓Large, well-supported ecosystem (many hosts, tooling, fine-tunes)
Best for developers who...
When to choose it (and when not to)
Reach for DeepSeek V3 when...
- →You maintain an existing V3 fine-tune or pinned deployment
- →You need to self-host or fine-tune without licensing friction
- →You want reproducible open-weight behavior pinned to a known version
Look elsewhere if...
- ✕You need current performance or a 1M context (use DeepSeek V4 Pro or V4.1 Flash)
- ✕You call DeepSeek's own API: the deepseek-chat name was retired on July 24, 2026
- ✕You need multimodal input
- ✕You need very long outputs beyond typical limits
How to use it
- ›Provide clear, explicit instructions - V3 follows structured prompts well
- ›Use system prompts to set role and output format (e.g. JSON)
- ›For coding, include full context and desired language/framework explicitly
- ›Pin to a specific hosted V3 build (e.g. the 0324 revision) for reproducibility
Quickstart
Pythonfrom openai import OpenAI
# Example via OpenRouter (third-party host of the open weights)
client = OpenAI(
api_key="YOUR_OPENROUTER_API_KEY",
base_url="https://openrouter.ai/api/v1",
)
resp = client.chat.completions.create(
model="deepseek/deepseek-chat",
messages=[{"role": "user", "content": "Write a Python function to parse an ISO date."}],
)
print(resp.choices[0].message.content)The open weights (deepseek-ai/DeepSeek-V3) can also be run locally with Ollama or vLLM instead of a hosted API.
API model id: deepseek-chat
Benchmarks
| Benchmark | Score | Notes |
|---|---|---|
| Pre-training scale | ~15T tokens | DeepSeek-V3 was pre-trained on nearly 15 trillion tokens (per model documentation). |
Compare DeepSeek V3
Compare DeepSeek V3 with any other model
Build a comparison →All model comparisons →