Seed-1.6 (AdaCoT)

General-purpose multimodal model series introducing Adaptive Chain-of-Thought (AdaCoT) — a reward-shaped RL recipe that teaches the model to decide whether to think before answering, instead of always thinking (like a reasoning model) or never thinking (like a base chat model). Three published variants:

• Seed-1.6 (AdaCoT): model auto-decides per-prompt; matches Full-CoT effectiveness while compressing CoT length significantly
• Seed-1.6-Thinking: always thinks; reasoning-mode peer of Seed1.5-Thinking with VLM capabilities
• Seed-1.6-FullCoT / NoCoT: ablations used to measure the AdaCoT trade-off; FullCoT triggers thinking ~90–100% of the time on AIME / BeyondAIME and matches Thinking-mode quality

256K context, multimodal (text + image), GUI-interaction-trained. Served via the Volcengine API. Not currently scored on Artificial Analysis.