State-of-the-art open-weight reasoning model. 680B-parameter MoE. Matches or exceeds GPT-4o on math, coding, and chain-of-thought — at roughly 5% of the price. Available via NovAI's zero-platform-fee Hong Kong gateway.
DeepSeek-V4-Pro is the flagship release of DeepSeek AI's fourth-generation Mixture-of-Experts architecture, launched January 2026. It is a 680-billion-parameter sparse MoE transformer with approximately 37B active parameters per token, trained on 14.8 trillion carefully curated multilingual tokens with a strong bias toward reasoning-heavy data — mathematical proofs, competitive-programming solutions, chain-of-thought corpora, and structured scientific reasoning.
What makes V4-Pro extraordinary is its economic profile. By publishing the weights open-source and shipping an aggressive MoE routing scheme, DeepSeek has compressed flagship-quality reasoning into a price point that is 15× cheaper than GPT-4o on output tokens. For reasoning-first workloads — coding assistants, math tutors, agent planners, structured-output pipelines — it is arguably the best price-performance large language model available in 2026.
On NovAI, V4-Pro is accessible through an OpenAI-compatible endpoint, streamed via our Hong Kong point-of-presence. You pay only the token rate; there is no subscription, platform markup, or seat fee. The companion DeepSeek-V4-Flash is available for high-throughput cheap-as-chips workloads at roughly a third of the Pro rate.
| Property | Value |
|---|---|
| Model ID | deepseek-v4-pro |
| Architecture | Sparse Mixture-of-Experts transformer (V4 routing) |
| Total parameters | ~680B |
| Active parameters / token | ~37B |
| Number of experts | 256 routed + 1 shared, top-8 routing |
| Context window (input) | 131,072 tokens (128K) |
| Max output tokens | 8,192 |
| Tokenizer | DeepSeek BPE (~100K vocabulary) |
| Training cutoff | October 2025 |
| Release date | January 2026 |
| Open weights | ✓ MIT-style license on Hugging Face |
| Modalities | Text in, text out |
| Streaming | ✓ Server-Sent Events |
| Function / tool calling | ✓ Parallel tool calls, OpenAI-compatible |
| JSON mode | ✓ Structured output via response_format |
| Temperature range | 0.0 – 2.0 (default 1.0 — lower recommended for reasoning) |
| Rate limit (default) | 60 RPM · 200K TPM |
| Rate limit (Scale tier) | 600 RPM · 2.5M TPM (unlock at $50 balance) |
| SLA | 99.9% monthly uptime |
| Direction | Per 1M tokens | Per 1K tokens |
|---|---|---|
| Input | $0.57 | $0.00063 |
| Output | $1.15 | $0.00190 |
| Model | Output price | vs DeepSeek-V4-Pro |
|---|---|---|
| OpenAI o1-preview | $60.00 | 69× more expensive |
| GPT 3.5 Sonnet | $15.00 | 17× more expensive |
| GPT-4o | $15.00 | 17× more expensive |
| Gemini 1.5 Pro | $10.50 | 12× more expensive |
| Doubao-Seed-2.0-Pro | $4.00 | 4.6× more expensive |
| Qwen3.6-Max | $4.80 | 5.5× more expensive |
| DeepSeek-V4-Pro (NovAI) | $1.15 | baseline |
| DeepSeek-V4-Flash (NovAI) | $0.57 | 0.33× (cheaper sibling) |
The same workload on GPT-4o would cost $750 → $7,500 → $37,500/month respectively — DeepSeek-V4-Pro delivers comparable output quality at 2–4% of the cost.
V4-Pro's 90.9 HumanEval and 65.2 LiveCodeBench put it in the GPT 3.5 Sonnet tier. Pair with Cursor, Cline, or Continue and you get flagship autocomplete / refactor / chat at a fraction of the monthly cost — most indie devs spend under $5/month of tokens.
AIME 2025: 79.8 and MATH: 92.4 make V4-Pro one of the best models publicly available for step-by-step mathematical work, theorem proving hints, and scientific derivation.
Reliable parallel tool calls plus low hallucination on long tool-use traces. A common cost-effective pattern: V4-Pro as planner + V4-Flash as worker, cutting total cost by another 40–60% without quality loss.
128K handles most monorepo backends, 90% of research papers, and full-novel-length legal contracts. Combine with a diff prompt for PR review workflows.
JSON mode + tight temperature gives near-deterministic schema extraction from unstructured text. Benchmarked at 98.4% valid-JSON rate over 10K production samples.
At around $1.15 per 1M output tokens, generating millions of synthetic training examples is finally affordable. Popular for fine-tuning smaller task-specific models.
curl https://aiapi-pro.com/v1/chat/completions \
-H "Authorization: Bearer $NOVAI_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "deepseek-v4-pro",
"messages": [
{"role":"system","content":"You are a rigorous reasoning assistant."},
{"role":"user","content":"Prove that sqrt(2) is irrational."}
],
"temperature": 0.2,
"max_tokens": 800
}'
from openai import OpenAI
client = OpenAI(
base_url="https://aiapi-pro.com/v1",
api_key="YOUR_NOVAI_API_KEY",
)
resp = client.chat.completions.create(
model="deepseek-v4-pro",
messages=[
{"role":"system","content":"You are a senior Rust engineer."},
{"role":"user", "content":"Write a lock-free MPSC queue in idiomatic Rust."},
],
temperature=0.2,
max_tokens=1200,
)
print(resp.choices[0].message.content)
print("Usage:", resp.usage)
stream = client.chat.completions.create(
model="deepseek-v4-pro",
messages=[{"role":"user","content":"Solve AIME 2024 problem 10 step by step."}],
stream=True,
)
for chunk in stream:
delta = chunk.choices[0].delta.content
if delta:
print(delta, end="", flush=True)
tools = [
{"type":"function","function":{
"name":"search_docs",
"description":"Search the internal docs",
"parameters":{"type":"object","properties":{"query":{"type":"string"}},"required":["query"]}}},
{"type":"function","function":{
"name":"run_sql",
"description":"Execute a read-only SQL query",
"parameters":{"type":"object","properties":{"sql":{"type":"string"}},"required":["sql"]}}},
]
resp = client.chat.completions.create(
model="deepseek-v4-pro",
messages=[{"role":"user","content":"How many MRR dollars did Asia contribute in Q1 2026?"}],
tools=tools,
tool_choice="auto",
parallel_tool_calls=True,
)
for call in resp.choices[0].message.tool_calls:
print(call.function.name, call.function.arguments)
import OpenAI from "openai";
const client = new OpenAI({
baseURL: "https://aiapi-pro.com/v1",
apiKey: process.env.NOVAI_API_KEY,
});
const resp = await client.chat.completions.create({
model: "deepseek-v4-pro",
messages: [{ role: "user", content: "Explain B-trees vs LSM-trees for an SRE interview." }],
temperature: 0.3,
});
console.log(resp.choices[0].message.content);
from langchain_openai import ChatOpenAI
llm = ChatOpenAI(
model="deepseek-v4-pro",
openai_api_base="https://aiapi-pro.com/v1",
openai_api_key="YOUR_NOVAI_API_KEY",
temperature=0.3,
)
print(llm.invoke("Summarize the pros and cons of CRDTs in 3 bullets").content)
Full documentation: aiapi-pro.com/#docs · docs.aiapi-pro.com
| Metric | P50 | P95 | P99 |
|---|---|---|---|
| TTFT (Hong Kong / CN / SEA) | 320ms | 460ms | 720ms |
| TTFT (Tokyo / Seoul) | 520ms | 740ms | 1.1s |
| TTFT (US West) | 900ms | 1.3s | 2.0s |
| TTFT (Europe) | 1.15s | 1.7s | 2.4s |
| Throughput (tokens/sec) | 58 | 42 | 31 |
| Long-context (>64K) TTFT | 1.8s | 2.9s | 4.7s |
Measurements aggregated from 30 days of production traffic. SLA: 99.9% monthly uptime. If DeepSeek upstream degrades, NovAI provides automatic fallback to V4-Flash with status-page notification.
| Model | Best for | Context | Input / 1M | Output / 1M |
|---|---|---|---|---|
| DeepSeek-V4-Pro | Flagship reasoning, coding, math, agents | 128K | $0.57 | $1.15 |
| DeepSeek-V4-Flash | High-throughput low-latency tasks: classification, summarization, simple chat | 128K | $0.08 | $0.17 |
Decision rule:
DeepSeek AI's flagship 680B-parameter MoE model released January 2026. ~37B active per token. Reasoning-first training pipeline. Matches GPT-4o on reasoning benchmarks at ~5% of the price.
$0.57/1M input, $1.15/1M output on NovAI. Zero platform fee. Credits never expire.
Yes — released under an open MIT-style license on Hugging Face. Self-host is possible; NovAI's API is the convenient hosted alternative.
128K input tokens. Max output per response: 8,192 tokens.
Yes — Server-Sent Events streaming and OpenAI-compatible tool calling including parallel invocation.
V4-Flash is ~4× cheaper and 2× faster, trained for throughput-sensitive workloads. Use Pro when quality matters, Flash when latency/cost matter. A hybrid planner(Pro)+worker(Flash) pattern typically wins both.
Ties or wins on reasoning benchmarks (MATH, GPQA, LiveCodeBench, AIME). Slightly behind on creative writing and general world knowledge. ~17× cheaper per output token.
No. NovAI does not retain prompts or completions. Only aggregate billing metadata (timestamp, token counts) is logged.
P50 TTFT 320ms from Hong Kong / China / SEA, 900ms from US West. Global throughput averages 45–70 tokens/sec.
Change two values in your OpenAI SDK code: base_url="https://aiapi-pro.com/v1" and model="deepseek-v4-pro". No other code changes required.
Zero platform fee. Credits never expire. OpenAI-compatible — no code changes needed. $2.00 free credit on signup.
Sign Up Free Compare All ModelsRelated: DeepSeek-V4-Flash · Doubao-Seed-2.0-Pro · Qwen3.6-Max · GLM-5.1