By name: bytedance-seed-2.0 description: | ByteDance Seed 2.0 Full Ecosystem skill for multi-modal AI routing. Use when: routing tasks to budget frontier LLMs, generating video/audio, selecting cost-effective models, or integrating Volcano Engine APIs. role: Specialist Executor intent: Route tasks to ByteDance's Seed ecosystem (LLM, video gen, TTS, music) for budget frontier reasoning and media generation. kpis: - model_routing_accuracy - cost_per_task - video_generation_success_rate status: active priority: high triggers: - task requires budget frontier reasoning model - video generation requested - TTS or music generation needed - cost optimization for AI inference execution: sequential dependencies: - VOLCANO_ENGINE_API_KEY
ByteDance Seed 2.0 Full Ecosystem
Overview
ByteDance's Seed ecosystem provides a comprehensive suite of AI models spanning LLM reasoning, video generation, text-to-speech, and music generation. These models offer frontier-competitive performance at significantly lower cost, making them the default choice for budget-sensitive tasks within A.I.M.S.
The ecosystem is accessed through Volcano Engine APIs and integrated via the UEF Gateway's model router.
Model Lineup
| Model | Category | Strengths | Cost Tier |
|---|---|---|---|
| Seed 2.0 LLM | Text/Reasoning | Frontier reasoning, 128K context | Low |
| Seed 2.0 Think | CoT Reasoning | Chain-of-thought, math, code | Low |
| Seaweed (Video) | Video Gen | Text-to-video, image-to-video | Medium |
| Seed-TTS | Speech | Zero-shot voice cloning, multilingual | Low |
| Seed Music | Audio | Text-to-music, style transfer | Low |
Pricing Advantage
Seed 2.0 LLM costs approximately 1/10th of GPT-4o and 1/5th of Claude Sonnet for equivalent reasoning tasks. This makes it the default routing target for:
- Summarization and extraction tasks
- Code generation (non-critical paths)
- Data analysis and transformation
- Bulk content generation
- Draft/iteration phases of FDH pipeline
Reserve higher-cost models (Claude, GPT-4o) for:
- Final output verification (ORACLE gates)
- Complex multi-step reasoning chains
- Safety-critical decisions
- User-facing conversational AI (ACHEEVY primary)
Routing Rules
- Default to Seed — All new tasks route to Seed 2.0 LLM unless the task explicitly requires a specific model's capabilities.
- Escalation Path — If Seed returns low-confidence output (self-reported or detected via ORACLE), escalate to Claude Sonnet, then Claude Opus.
- Media Tasks — Video generation always routes to Seaweed. TTS routes to Seed-TTS. Music routes to Seed Music.
- Context Window — Tasks exceeding 128K tokens must route to models with larger context (Gemini 2.5 Pro at 1M tokens).
- Latency Sensitive — Real-time chat responses use ACHEEVY's primary model (Claude). Seed is for async/batch tasks.
Anti-Patterns
- Never use Seed for safety-critical decisions (Chicken Hawk compliance, deployment authorization).
- Never expose Volcano Engine API keys to the client side.
- Never bypass the UEF Gateway model router to call Seed directly from frontend code.
- Never assume Seed output is final — all Seed outputs in production must pass through at least one verification gate.
Integration Notes
- All Seed API calls go through
backend/uef-gateway/via the OpenClaw router's model selection layer. - The
VOLCANO_ENGINE_API_KEYenvironment variable must be set in the UEF Gateway's environment. - Response format normalization happens in the gateway — downstream consumers receive a unified response schema regardless of which model was used.
- Rate limits: Seed 2.0 LLM supports up to 100 RPM on the standard tier. Video generation is limited to 10 concurrent jobs.
- For detailed API specifications, benchmark tables, ecosystem catalog, and use case examples, see
references/api-patterns.md.