SetupAI
EN
來源較早收集於 14m

字節跳動火山引擎推出 Agent 與 Coding Plan 限時優惠

字節跳動火山引擎推出 Agent 與 Coding Plan 限時優惠
PostLinkedIn
🏠閱讀原文： IT之家
#cloud-computing#model-as-a-service#pricing-strategyvolcengine-agent-plan-&-coding-planvolcenginebytedancedeepseekminimax

💡以極具競爭力的價格獲取 DeepSeek V4 與 MiniMax M3 等頂尖模型的使用權。

⚡ 30 秒速覽

有什麼變化

Coding Plan 與 Agent Plan 優惠活動持續至 2026 年 8 月 27 日。

為什麼重要

火山引擎的激進定價策略加劇了中國企業級 AI 市場的競爭，使開發者與初創企業能以更低成本獲取頂尖模型能力。

下一步行動

如果您正在尋找具成本效益的 AI 輔助開發環境，建議評估火山引擎的 Coding Plan 套餐。

誰應關注：Developers & AI Engineers

關鍵要點

  • •Coding Plan 與 Agent Plan 優惠活動持續至 2026 年 8 月 27 日。
  • •整合模型包括 MiniMax M3、DeepSeek V4 及 GLM-5.1 等。
  • •首兩個月最低價格降至 9.9 元人民幣（原價 40 元）。
  • •Agent Plan 內建字節自研的 Doubao-Seed 模型，支援多模態任務。

🧠 深度解析

背景與延伸：來自公開資料，非原文內容。引用 28 個來源。

🔑 增強重點摘要

  • •Volcengine's discount strategy is part of a broader push to capture a larger share of China's public cloud LLM market, where it already held a 46.4% market share by token calls as of mid-2025, surpassing Baidu AI Cloud and Alibaba Cloud combined.
  • •The integrated models offer advanced capabilities: MiniMax M3 features a 1-million-token context window and native multimodal understanding (image and video), excelling in coding and agentic tasks with a proprietary MiniMax Sparse Attention (MSA) architecture.
  • •DeepSeek V4, another featured model, is a 1-trillion-parameter Mixture-of-Experts (MoE) model with a 1-million-token context window, native multimodal support (text, images, video, audio), and is positioned as a cost-effective competitor to Western frontier models.
  • •GLM-5.1, also included, is a 754-billion parameter MoE model with a 202K context window, designed for long-horizon agentic engineering, demonstrating strong performance in coding and autonomous task execution over extended periods.
  • •The Agent Plan specifically leverages proprietary Doubao-Seed models, such as Doubao-Seed-2.0-lite, which offers full-modal understanding (video, images, audio, text), enhanced agentic capabilities with improved multi-turn instruction compliance, and integrated GUI understanding and execution.
📊 競品分析▸ Show

markdown

Feature/ProviderByteDance Volcengine (Agent/Coding Plan)Baidu AI Cloud (DuClaw)Tencent Cloud (Agent Development Platform)
Pricing (Base)40 RMB/month (Agent Plan Basic), 9.9 RMB/month (discounted for 2 months)17.8 RMB/month (introductory rate)Varies by tier (Free, Starter, Team, Enterprise); Hunyuan model prices increased over 450% in March 2026
Key ModelsMiniMax M3, DeepSeek V4, GLM-5.1, Doubao-Seed modelsOpenClaw agent platformHunyuan series, GLM 5, MiniMax 2.5, Kimi 2.5 (previously)
Core OfferingIntegrated Agent Plan with multimodal models, web search, Vision Embedding, AFP billingZero-deployment AI agent service for OpenClawAgent development and management platform with NLP, multi-language, dialog management, and RAG+LLM capabilities
Context WindowUp to 1M tokens (MiniMax M3, DeepSeek V4), 202K tokens (GLM-5.1), 256K tokens (Doubao-Seed)Not explicitly detailed for DuClaw, but OpenClaw compatible models support large contextsVaries by model, e.g., Doubao-seed-1.6 supports 256K contexts
MultimodalityNative multimodal (MiniMax M3, DeepSeek V4, Doubao-Seed-2.0-lite)Not explicitly detailed for DuClaw, but OpenClaw can integrate visionSupports multi-modal understanding (e.g., Doubao-seed-1.6)
BenchmarksMiniMax M3: 59.0% on SWE-Bench Pro; DeepSeek V4 Pro: ~91.2% on SWE-Bench Verified; GLM-5.1: 58.4% on SWE-Bench ProNot specified for DuClaw directly, relies on underlying modelsNot explicitly detailed for ADP, but Hunyuan models have benchmarks

🛠️ 技術深入

  • MiniMax M3: Employs a proprietary MiniMax Sparse Attention (MSA) architecture, which reduces per-token compute to one-twentieth of the prior generation at 1-million-token context length, enabling over 9x faster prefill and 15x faster decoding. It is natively multimodal, trained from scratch on interleaved text, image, and video data.
  • DeepSeek V4: Features a Mixture-of-Experts (MoE) architecture with 1.6 trillion total parameters and approximately 49 billion active parameters per token. It utilizes a hybrid attention architecture combining Compressed Sparse Attention (CSA) and Heavily Compressed Attention (HCA) to achieve significant efficiency gains (27% of single-token inference FLOPs and 10% of KV cache compared to V3.2 at 1M-token context). It supports native multimodal input (text, images, video, audio) and offers three reasoning modes: Non-think, Think High, and Think Max.
  • GLM-5.1: A 754-billion parameter Mixture-of-Experts (MoE) model with 40 billion active parameters per token and a 202,752-token context window. Its core innovation is an iterative reasoning framework that allows the model to sustain optimization over hundreds of rounds, enabling autonomous work on complex tasks for up to 8 hours by continuously revisiting reasoning, running experiments, and revising strategies.
  • Doubao-Seed Models: The Doubao-Seed-2.0-lite is a full-modal understanding model capable of native unified understanding of video, images, audio, and text. It features upgraded Agent and Coding capabilities, improved compliance with multi-turn complex instructions, stronger self-decomposition and verification, and integrated GUI understanding and execution, allowing it to perform operations like clicking and dragging on interfaces.

🔮 前景展望基於引用來源的 AI 分析

ByteDance's aggressive pricing strategy for AI services will intensify the price war in China's cloud AI market.
By offering significant discounts and bundling capabilities, Volcengine aims to rapidly acquire developers and market share, forcing competitors like Alibaba Cloud and Tencent Cloud to respond with similar pricing adjustments.
The focus on integrated Agent and Coding plans signifies a shift towards more autonomous and task-oriented AI applications.
Volcengine's bundling of models with 'Harness' tools like web search and vision embedding, along with the advanced agentic capabilities of models like MiniMax M3, DeepSeek V4, GLM-5.1, and Doubao-Seed, indicates a move beyond simple language generation to complex, multi-step task execution.
ByteDance is strategically positioning Volcengine to become a dominant player in enterprise AI, leveraging its consumer AI expertise.
By commercializing its internal AI stack, including the Doubao model family and the HiAgent platform, and aggressively targeting enterprise clients with competitive pricing and comprehensive solutions, ByteDance seeks to diversify its revenue beyond social media and challenge established cloud incumbents.

⏳ 時間線

2012
ByteDance was established.
2021
Volcengine, ByteDance's cloud computing and AI service platform, was launched.
2023-08
ByteDance launched its chatbot app Doubao (initially named Skylark).
2024-05
ByteDance first released the Doubao model family.
2025-07-30
Volcengine launched the upgraded Doubao Large Model 1.6 series and other new offerings, doubling down on open-source for AI agents.
2026-05-11
Volcengine officially launched the 'Agent Plan' subscription package, integrating multimodal models and Harness tools.
🏠閱讀原文： IT之家
📰

AI 週報

閱讀本週精選 AI 大事摘要 →

👉相關動態

同主題

探索 #cloud-computing

同產品

更多關於 volcengine-agent-plan-&-coding-plan

同來源

最新來自 IT之家

AI 策展新聞聚合。所有內容版權歸原始發布者所有。
原始來源： IT之家 ↗

這是摘要，不是原文。去看原站，或訂閱每週簡報。

每週電子報

每週一封，可隨時退訂。