Doubao AI

Just Now: Doubao Launches Multiple Large Models — Making DeepSeek Easier to Use, Audio Sora Wows the Crowd

Honghao Wang

— 4 min read
Just Now: Doubao Launches Multiple Large Models — Making DeepSeek Easier to Use, Audio Sora Wows the Crowd

AI in 2025: Performance vs. Cost — The Ongoing Dilemma

Frankly speaking — it’s already 2025, yet asking AI questions still feels polarized:

  • Simple queries: Responses come instantly, but sometimes feel incomplete.
  • Complex queries: Requires deep reasoning, taking over 30 seconds to respond.
  • Token burn: Every reply consumes tokens, directly impacting business costs.

For enterprises, token consumption = cost. Deep reasoning improves results, but increases latency, consumption, and expenses — a persistent, industry-wide pain point.

image

---

Token Usage Explosion

AI adoption is accelerating:

  • Doubao Large Model daily calls exceeded 30 trillion tokens by Sept 2024, up 80% since May.
  • This is just one segment of the global market.

Companies face a trade-off:

  • Lightweight models: Save money, sacrifice quality.
  • Top-tier models: Deliver quality, incur high costs.

---

Volcano Engine’s Answer: Balancing Quality and Cost

October 16 — FORCE LINK AI Innovation Roadshow, Wuhan:

Volcano Engine unveiled four new products:

  • Doubao Large Model 1.6 — Native support for 4 adjustable thinking lengths.
  • Doubao 1.6 Lite — Halves costs, improves results.
  • Doubao Speech Synthesis Model 2.0.
  • Voice Cloning Model 2.0.

📊 IDC Report: 1H 2025 — Volcano Engine captured 49.2% of China’s public cloud LLM service market.

Translation: 1 in 2 companies using cloud-based large models is on Volcano Engine.

image

---

Tackling Low Adoption of Deep Reasoning

At launch, one insight stood out:

  • Deep reasoning mode boosts results by 31%,
  • But due to latency, cost, and token usage, adoption is only 18%.
  • In short: Businesses want it, but can’t afford it.

---

Doubao Large Model 1.6: "Tiered Thinking Length"

First domestic model with native tiered reasoning length adjustment:

  • Minimal — Best for lookups, saves tokens.
  • Low — Balanced quality & cost.
  • Medium — Higher quality with moderate cost.
  • High — Maximum depth for complex logic.

Analogy: Like adding a gearbox to AI — choose the gear for the task.

Low-tier example:

  • Output tokens: ↓ 77.5%
  • Thinking time: ↓ 84.6%
  • Result quality: No change

---

Doubao 1.6 Lite: Better Speed, Lower Cost

  • 14% improvement vs. Doubao 1.5 Pro in enterprise testing.
  • 53.3% lower cost (0–32k input range).
  • Higher unit token value density = More value per cost unit.
image

---

Upgrades in Speech Models

Doubao Speech Synthesis Model 2.0 & Voice Cloning Model 2.0:

  • Enhanced emotional expression & instruction compliance.
  • Accurate complex formula reading — Essential for education.
  • Market average: < 50% accuracy.
  • Doubao new models: Up to 90% accuracy in math/science formula reading (primary–high school level).
image

---

Under the Hood

Built on Doubao LLM architecture for:

  • Semantic understanding before speaking text.
  • Precise emotional control via natural language commands:
  • “Make it gentler.”
  • “Read it with energy.”

---

Real-World Application Example

Children’s Picture Book — Protecting the Baer’s Pochard in Wuhan

Tools used:

  • Beanbag Image Creation Model Seedream4.0 — Illustrations.
  • Beanbag Speech Synthesis Model 2.0 — Emotional narration.

Beanbag Voice Models now span seven domains:

Speech synthesis, recognition, voice cloning, real-time speech, simultaneous interpretation, music creation, and podcast production — integrated into 460M+ intelligent devices.

image

---

Tan Dai identified:

  • Deep-thinking models integrating multi-modal understanding.
  • Video, image, and voice models approaching production grade.
  • Enterprise-level Agents maturing.

---

Choosing the Right Model: Smart Model Router

Volcano Engine launched China’s first intelligent model selection system:

Available on Volcano Ark, with three routing modes:

  • Balanced Mode
  • Quality Priority Mode
  • Cost Priority Mode
image

---

Why It Matters

Tasks differ in value density per token:

  • Simple customer service → lightweight model.
  • Complex medical analysis → top-tier model.

Smart routing lets AI decide token burn per task, supporting:

  • Beanbag LLM
  • DeepSeek
  • Qwen
  • Kimi

---

DeepSeek Routing Results

  • Quality Priority: +14% performance vs. direct DeepSeek-V3.1.
  • Cost Priority: Similar quality vs. DeepSeek-V3.1, cost ↓ 70%+.

---

The Positive Feedback Loop

How smart routing fuels the AI industry:

> Better models → more applications → higher token usage → smarter routing → lower unit cost → demand grows → cycle continues.

Tokens are becoming the kilowatt-hour of the AI era — measurable, billable, trackable.

  • OpenAI joined the “Trillion Token Club.”
  • Google burns 1.3 trillion tokens monthly.
image

---

From Good Models to Good Experiences

The end goal:

  • Simple queries: Fast, accurate.
  • Complex problems: Deep, efficient reasoning.
  • Automatic model choice with Smart Model Router.
  • Natural language voice control without technical overhead.

---

Content Creation Meets AI Efficiency

Platforms like AiToEarn官网 connect:

  • AI model selection
  • Content generation
  • Cross-platform publishing
  • Analytics and model rankings (AI模型排名)

Supported platforms:

  • Douyin, Kwai, WeChat, Bilibili, Rednote (Xiaohongshu), Facebook, Instagram, LinkedIn, Threads, YouTube, Pinterest, X (Twitter).

Goal: Monetize AI creativity globally — efficiently.

---

Final Takeaway

The future of AI is about powerful models, intelligent routing, and integrated ecosystems that turn ideas into experiences everywhere.

Doing more with less — and doing it smarter.

---

Would you like me to also create an infographic-style summary table of these AI products and their improvements so this Markdown becomes more visually concise? That could make the differences between each Doubao model instantly clear.

Read more