Seed by ByteDance ✓

Multimodal input handling, supporting text, image, audio, and video for generation workflows.

Multi-shot video generation with better scene continuity instead of single-clip outputs.

Native audio generation integrated into video outputs.

Strong reference control, allowing subject, style, and scene consistency across shots.

Supports editing and extension workflows like continuing or modifying existing videos.

Stronger emphasis on real-world complexity evaluation via a 4-part framework (Science Discovery, Vibe Coding, Context Learning, Real-World Tasks) instead of Seed1.8’s broader benchmark grouping.

Deeper GUI-agent focus with explicit end-to-end evaluations in heavy “real app” environments like FreeCAD (CAD) and CapCut (video editing), which are not used as named GUI testbeds in Seed1.8.

More direct focus on reducing visual hallucinations and improving structured extraction from screenshots, charts, and scanned documents compared to Seed1.8’s more general multimodal capability framing.

Tool orchestration is treated as a more central capability axis, highlighting orchestration benchmarks (for example MCP-Mark) beyond the tool-use framing in Seed1.8.

The write-up shifts from “generalized real-world agency” toward “intelligence frontier for real-world complexity,” putting more weight on long-horizon, high-value workflows (research, coding projects, context learning) as the organizing target.

Described as a general-purpose agent model rather than a plain LLM.

Emphasizes completing tasks in complex workflows efficiently and accurately.

Positioned as the most “agentic” entry among the four.

Best fit for tool-like, multi-step task execution scenarios.

Framed explicitly around workflow completion on the Models page.

Presented as a newer iteration after Seed1.5 on the Models page.

Emphasizes integrated multimodal capabilities.

Highlights “adaptive thinking” to balance performance with reasoning.

Positioned between Seed1.5 and Seed1.8 in the lineup.

Best fit when you want multimodal plus a stronger reasoning tradeoff knob.

Vision-language model focused on multimodal understanding and generation.

Positioned as state of the art on 38 out of 60 public benchmarks.

Listed separately from the general Seed1.5 LLM line.

Best fit when inputs include images plus text, not just text.

Framed as a benchmark-leading multimodal model on the Models page.

General-purpose LLM positioned for knowledge, coding, and reasoning.

Uses a sparse MoE architecture per the Doubao-1.5-pro page.

Framed as a balance between model quality and inference efficiency.

Best fit for text-first tasks like coding and general reasoning.

Has its own dedicated technical page dated 2025-01-22.