Seed1.6 (Doubao-Seed-1.6)

What it is

Seed1.6 is the general-purpose large model series from ByteDance Seed, announced as Doubao Large Model 1.6 at Volcano Engine's FORCE conference on 11 June 2025.^[1] It accepts multimodal input, is built to interact with graphical user interfaces, and carries a 256K context window; developers reach it through Volcano Engine's API.^[2] Three variants are published: Seed1.6, Seed1.6-Thinking, and Seed1.6 (Adaptive CoT).^[2]

How it is built

The model is a sparse mixture-of-experts (MoE) network with 230 billion parameters in total and 23 billion active.^[2] In a sparse MoE, each token is routed to a small subset of the network's experts instead of passing through all the weights, so the compute spent per token stays far below what the total parameter count would suggest.

Pre-training ran in two stages. The first stage used web pages, books, papers, code and other data; a second, multimodal stage added visual data alongside text.^[2] Long-context training then extended the maximum sequence length from 32K to 256K tokens.^[2]

Thinking or not thinking

The series' distinguishing feature is that reasoning is decided per request rather than fixed by the model you call. Seed1.6 (Adaptive CoT) uses AdaCoT, which looks at each prompt and chooses whether to run an extended chain of thought before answering; Seed1.6-Thinking always thinks (FullCoT), and the plain Seed1.6 never does (NoCoT).^[2]

ByteDance Seed reported how often AdaCoT chose to think: about 37% of MMLU prompts, 70% of MMLU-Pro prompts, and 90–100% of AIME and BeyondAIME problems.^[2] The pattern is that a general knowledge question usually gets a direct answer, while competition mathematics nearly always triggers reasoning.

Results

On China's 2025 gaokao, using the Shandong papers, Seed1.6-Thinking scored 683 in the humanities track and 676 in the science track after a re-test with images.^[2] The science score had first come out at 648; supplying higher-resolution images in the re-test raised it to 676.^[2] The model also scored 100% on JEE Advanced mathematics across five samplings.^[2]

Price and access

Volcano Engine priced the model by input length, with a single rate covering thinking, multimodal and plain-text use: 0.8 yuan per million input tokens and 8 yuan per million output tokens in the 0–32K tier.^[1] The company said its overall cost was about one third of Doubao 1.5 deep-thinking or DeepSeek R1.^[1]

Developers call Seed1.6 and Seed1.6-Thinking through Volcano Engine's open API under the model IDs doubao-seed-1-6-250615 and doubao-seed-1-6-thinking-250615.^[2]

Where it sits

Seed1.6 brings three things into one series: multimodal input and a 256K context window, a switch that decides per prompt whether to reason, and a price that does not change between thinking and non-thinking, multimodal and plain-text calls.^[1][2] It is distributed through Volcano Engine, the cloud platform through which ByteDance's models are offered to outside developers.^[1][2]