DeepSeek V4 Pro
DeepSeek’s flagship open-weight model, built for coding, reasoning, and agentic work — served on the Arcware Inference Stack at native weights with zero quantization.
Model details
DeepSeek V4 Pro is DeepSeek’s flagship Mixture-of-Experts model and the current state of the art among open-source models, leading on coding, reasoning, and agentic benchmarks. Of its 1.6 trillion parameters, 49 billion are active per token, and it handles context windows up to 1 million tokens.
Its hybrid attention design pairs Compressed Sparse Attention (CSA) with Heavily Compressed Attention (HCA), which is what makes very long contexts economical: at the full 1M-token window it needs only 27% of the single-token inference FLOPs and 10% of the KV cache of DeepSeek V3.2.
The model was pre-trained on more than 32 trillion tokens, then refined with a two-stage post-training run combining supervised fine-tuning, reinforcement learning, and on-policy distillation. The weights are fully open under the MIT license — and on Arcware they are served exactly as released, with zero quantization.
Arcware pricing per 1M tokens
50% below the next lowest price on the market.
OpenAI-compatible
Point your existing SDK at Arcware and stream.
from openai import OpenAI
client = OpenAI(
base_url="https://api.arcware.us/v1",
api_key=os.environ["ARCWARE_API_KEY"],
)
stream = client.chat.completions.create(
model="deepseek/deepseek-v4-pro",
messages=[{"role": "user", "content": "Plan a database migration."}],
stream=True,
)
for chunk in stream:
print(chunk.choices[0].delta.content or "", end="")