671B
Parameters
MoE (37B active)
AIVory Smart Inference · DeepSeek
Open WeightDeepSeek V3.2 - 671B brain, 37B price tag.
The largest open-weight MoE model in production. 671B total parameters, 37B active per token, 160K context. Top scores on code and math benchmarks.
At a glance
DeepSeek V3.2 via Smart Inference.
Routed to the cheapest live provider on every request. Same model, lower bill.
160K
Context window
tokens per request
—
Cheapest input price
per 1M tokens via Smart Inference
—%
Cheaper
vs the priciest endpoint for the same model
Pricing
DeepSeek V3.2: price comparison.
Our cheapest routed price vs what you'd pay elsewhere.
Prices in USD per 1M input tokens · updated live from the router
Quick start
Use DeepSeek V3.2 in three lines.
OpenAI-compatible. Swap your base_url, keep everything else.
from openai import OpenAI
client = OpenAI(
base_url="https://smart.aivory.net/v1",
api_key="sk-aivory-...",
)
resp = client.chat.completions.create(
model="deepseek-v3.2",
messages=[{"role": "user", "content": "Ship it."}],
)
print(resp.choices[0].message.content)import OpenAI from "openai";
const client = new OpenAI({
baseURL: "https://smart.aivory.net/v1",
apiKey: process.env.AIVORY_API_KEY,
});
const resp = await client.chat.completions.create({
model: "deepseek-v3.2",
messages: [{ role: "user", content: "Ship it." }],
});
console.log(resp.choices[0].message.content);curl https://smart.aivory.net/v1/chat/completions \
-H "Authorization: Bearer $AIVORY_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "deepseek-v3.2",
"messages": [{"role": "user", "content": "Ship it."}]
}'That's the whole migration. Get early access and try it →
Use cases
What DeepSeek V3.2 is great at.
-
Code Review and Generation
V3.2 scores within a few points of GPT-4o on HumanEval and MBPP. Feed it a pull request diff with 10K tokens of context and get line-by-line review with suggested fixes. Cost per review: under $0.01.
-
Mathematical and Scientific Computation
Strongest open model on MATH and GSM8K benchmarks. Use it for financial modeling, symbolic derivation, or step-by-step proofs. The 37B active parameter count keeps per-token cost close to a 70B dense model.
-
160K Context Window
The largest context window in the Smart Inference catalogue. Summarize entire repositories, analyze multi-document research sets, or maintain conversation history across hundreds of turns without chunking.
FAQ
Common questions.
How does DeepSeek V3.2 pricing work on AIVory?
You pay per token at whatever rate the cheapest live provider charges. The MoE architecture means V3.2 inference costs roughly what a 70B dense model costs, despite having 671B total parameters. Check the live table above for current rates.
DeepSeek V3.2 vs GPT-4o - how do they compare?
V3.2 matches or beats GPT-4o on code (HumanEval) and math (MATH, GSM8K) benchmarks. GPT-4o is stronger on creative writing and some reasoning tasks. V3.2 costs a fraction of GPT-4o's price.
Why does it say 671B if only 37B are active?
Mixture-of-experts (MoE) models have many expert sub-networks but only activate a few per token. V3.2 has 671B total parameters split across experts, but each token only passes through 37B. You get big-model quality at small-model inference cost.
Is V3.2 good for non-English languages?
Yes. DeepSeek trained V3.2 on a broad multilingual corpus with strong Chinese, Japanese, Korean, and European language coverage. It's one of the best open models for CJK tasks.
Start using DeepSeek V3.2. Today.
Swap your base_url, keep shipping.
The router is live. Credits from $10, no subscription, no commitment.