ByteDance's flagship 3.2-trillion-parameter MoE model. GPT-4-class reasoning, 256K context, state-of-the-art Chinese understanding — via NovAI's zero-platform-fee Hong Kong gateway with an exclusive 10% distributor discount.
Doubao-Seed-2.0-Pro is the flagship closed-source model in ByteDance's Doubao Seed family, released in April 2026. It is a sparse Mixture-of-Experts (MoE) transformer with 3.2 trillion total parameters and approximately 280B active parameters per forward pass, trained on a 15T-token multilingual corpus with heavy emphasis on Chinese-language and bilingual reasoning data.
It occupies the top-tier of ByteDance's lineup above Doubao-Seed-2.0-Lite (for high-throughput inference) and Doubao-Seed-2.0-Code (code-specialized). Where Lite is tuned for cost-per-token and Code for HumanEval-class tasks, Pro is the default choice when you want broad-capability reasoning on complex, open-ended, Chinese-first workloads.
On NovAI, Doubao-Seed-2.0-Pro is served through a direct distributor relationship with ByteDance Volcano Engine. We pass through the entire upstream capability surface — 256K context, streaming, function calling, JSON mode, vision inputs — with no schema changes required. Requests are routed via our Hong Kong SAR point-of-presence for sub-500ms first-token latency to Asian users and sub-second latency globally.
base_url and model only; no client-library migration. Works with OpenAI Python, Node.js, Go, LangChain, LlamaIndex, Vercel AI SDK, Cursor, Cline, and any tool that speaks Chat Completions.| Property | Value |
|---|---|
| Model ID | doubao-seed-2.0-pro |
| Architecture | Sparse Mixture-of-Experts (MoE) transformer |
| Total parameters | ~3.2T |
| Active parameters / token | ~280B |
| Context window (input) | 262,144 tokens (256K) |
| Max output tokens | 16,384 |
| Tokenizer | ByteDance BPE (shared Chinese/English/code vocabulary, ~151K tokens) |
| Training cutoff | December 2025 |
| Release date | April 2026 |
| Modalities | Text in, text out. Vision (image input) available. |
| 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) |
| Top-p range | 0.0 – 1.0 (default 0.7) |
| Rate limit (default) | 60 RPM · 150K TPM |
| Rate limit (Scale tier) | 600 RPM · 2M TPM (unlock at $50 balance) |
| SLA | 99.9% monthly uptime |
| Tier | Input (per 1M tokens) | Output (per 1M tokens) |
|---|---|---|
| Official Volcano Engine | $0.44 | $2.22 |
| NovAI (with 10% discount) | $0.40 | $2.00 |
| You save | 10% | 10% |
| Model | Output price | vs Doubao-Pro |
|---|---|---|
| GPT-4o (OpenAI) | $15.00 | 7.5× more expensive |
| GPT 3.5 Sonnet (Anthropic) | $15.00 | 7.5× more expensive |
| Gemini 1.5 Pro (Google) | $10.50 | 5.25× more expensive |
| Qwen3.6-Max (Alibaba) | $6.24 | 3.12× more expensive |
| Doubao-Seed-2.0-Pro (NovAI) | $2.00 | baseline |
| DeepSeek-V4-Pro (NovAI) | $0.40 | 0.2× (cheaper, weaker on Chinese) |
All prices shown in USD. Billing is prepaid — credits never expire. No hidden platform fee, no minimum top-up. First-time users get $2.00 free credit on signup, enough for roughly 1.25M input tokens or 125K output tokens.
Doubao's native Chinese excellence plus the 256K context window means you can stuff entire policy documents, HR handbooks, or research reports into a single call — no vectorstore chunking dance. Typical retrieval quality is 8–12 points higher than embedding-based hybrid search with GPT-4o.
256K is enough for a 50-file TypeScript backend. Point it at a whole pull request, ask for architectural critique, and it will trace state across files reliably. Pairs well with Cursor and Cline out of the box.
With parallel tool calls and JSON mode, Doubao-Pro runs complex 20+ step planner-executor loops with low error accumulation. Benchmarks on τ-bench and WebArena place it within 3 points of GPT 3.5 Sonnet at a quarter the output cost.
256K handles most 10-K filings, full legal contracts, or 300-page research monographs in one shot. Needle-in-a-haystack retrieval accuracy stays above 94% up to the 240K mark.
Native-feeling Chinese copywriting, culturally adapted bilingual translation, and tone-matched brand content. Outperforms GPT-4o on human-preference evaluations for xiaohongshu / WeChat-style content.
Strong at reading pasted CSV / TSV / markdown tables and reasoning over them. Use JSON mode with a typed schema for reliable structured-extraction pipelines.
curl https://aiapi-pro.com/v1/chat/completions \
-H "Authorization: Bearer $NOVAI_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "doubao-seed-2.0-pro",
"messages": [
{"role":"system","content":"You are a helpful assistant."},
{"role":"user","content":"Explain MoE architectures in 2 paragraphs."}
],
"temperature": 0.7,
"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="doubao-seed-2.0-pro",
messages=[
{"role": "system", "content": "You are a senior Python engineer."},
{"role": "user", "content": "Write a FastAPI health-check endpoint."},
],
temperature=0.3,
max_tokens=500,
)
print(resp.choices[0].message.content)
print("Usage:", resp.usage)
stream = client.chat.completions.create(
model="doubao-seed-2.0-pro",
messages=[{"role":"user","content":"Write a haiku about Hong Kong."}],
stream=True,
)
for chunk in stream:
delta = chunk.choices[0].delta.content
if delta:
print(delta, end="", flush=True)
tools = [{
"type": "function",
"function": {
"name": "get_weather",
"description": "Get current weather for a city",
"parameters": {
"type": "object",
"properties": {"city": {"type": "string"}},
"required": ["city"],
},
},
}]
resp = client.chat.completions.create(
model="doubao-seed-2.0-pro",
messages=[{"role":"user","content":"What's the weather in Hong Kong?"}],
tools=tools,
tool_choice="auto",
)
tool_call = resp.choices[0].message.tool_calls[0]
print(tool_call.function.name, tool_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: "doubao-seed-2.0-pro",
messages: [{ role: "user", content: "Top 3 SEO tips for a SaaS landing page." }],
temperature: 0.6,
});
console.log(resp.choices[0].message.content);
resp = client.chat.completions.create(
model="doubao-seed-2.0-pro",
messages=[
{"role":"system","content":"Extract a product object as JSON."},
{"role":"user", "content":"The new iPhone Mini is $799, releases Sep 2026, colors: red, blue."},
],
response_format={"type":"json_object"},
)
# Returns: {"name":"iPhone Mini","price_usd":799,"release":"2026-09","colors":["red","blue"]}
Full documentation: aiapi-pro.com/#docs · docs.aiapi-pro.com
| Metric | P50 | P95 | P99 |
|---|---|---|---|
| TTFT (Hong Kong / CN / SEA) | 280ms | 420ms | 690ms |
| TTFT (Tokyo / Seoul) | 450ms | 680ms | 1.0s |
| TTFT (US West) | 820ms | 1.2s | 1.9s |
| TTFT (Europe) | 1.10s | 1.6s | 2.3s |
| Throughput (tokens/sec) | 52 | 38 | 28 |
| Long-context (>128K) TTFT | 2.1s | 3.4s | 5.6s |
Measurements aggregated from 30 days of production traffic across >4,000 distinct customers. SLA commitment: 99.9% monthly uptime. If Volcano Engine upstream degrades, NovAI provides automatic fallback to Doubao-Seed-2.0-Lite with proactive status-page notification.
| Model | Best for | Context | Input / 1M | Output / 1M |
|---|---|---|---|---|
| Doubao-Seed-2.0-Pro | Flagship reasoning, Chinese-first RAG, complex agents | 256K | $0.40 | $2.00 |
| Doubao-Seed-2.0-Lite | High-throughput classification, summarization, cheap RAG | 128K | $0.075 | $0.45 |
| Doubao-Seed-2.0-Code | Code generation / completion / review at SOTA HumanEval | 128K | $0.15 | $1.00 |
Quick rule of thumb:
ByteDance's flagship Mixture-of-Experts large language model released in 2026. 3.2T total parameters, ~280B active per token, 256K context, GPT-4-class capability with native Chinese excellence.
On NovAI: $0.40/1M input, $2.00/1M output — a 10% discount vs official Volcano Engine pricing, zero platform fee, credits never expire.
Yes. Endpoint is https://aiapi-pro.com/v1/chat/completions. Drop-in replace base_url and model. Supports streaming, tool calls, JSON mode, parallel function calling.
256K input tokens (262,144). Max output per response: 16,384 tokens. Suitable for full-report QA, monorepo analysis, or 40+ round agent workflows.
Yes, set stream: true. First-token latency is 280–420ms from Hong Kong / China / SEA.
Yes — OpenAI-compatible tool calls with parallel invocation and structured JSON response.
Matches or beats GPT-4o on Chinese benchmarks. Trails GPT 3.5 Sonnet by ~2 points on HumanEval. ~75% cheaper per output token than both.
No. NovAI does not retain prompts or completions for training. Only billing metadata (timestamp, token count) is logged.
P50 TTFT 280ms from Hong Kong, 450ms from Tokyo, 820ms from US West. Global throughput averages 42–58 tokens/sec.
Sign up at aiapi-pro.com, receive $2.00 free credit, grab your API key, point your OpenAI SDK at our base URL. Takes under 90 seconds.
Zero platform fee. Credits never expire. OpenAI-compatible API — no code changes needed. $2.00 free credit on signup.
Sign Up Free Compare All ModelsRelated: Doubao-Seed-2.0-Lite · Doubao-Seed-2.0-Code · DeepSeek-V4-Pro · Qwen3.6-Max · GLM-5.1