Doubao-1.5-pro

What it is

Doubao-1.5-pro is a multimodal model from ByteDance, released on 22 January 2025 as a closed model. Multimodal means it can work with more than one kind of input, here text, images and speech. It is the successor to Doubao Pro, the high-end model of the Doubao family launched earlier.^[1][2]

Background

Doubao-1.5-pro was developed as a step up from the first Doubao models. Its architecture was designed as a sparse mixture-of-experts model, with the design chosen during pre-training and carried through to inference.^[1] A mixture-of-experts model splits its internal capacity into many specialist sub-networks and activates only some of them for each token. This lets a model hold a large amount of knowledge while using less computation for each word it produces.

Key features

ByteDance said that a mixture-of-experts model activating about one seventh of a dense model's parameters matched or beat that dense model when both were trained on the same 9 trillion tokens. The company described this as a 7x performance leverage, compared with an industry norm of under 3x.^[1] ByteDance did not disclose the total or the activated parameter count of the model.

The team also reported that its mixture-of-experts base model outperformed Llama 3.1-405B in pre-training evaluations, even though it activated far fewer parameters.^[1] ByteDance said that no data from other models was used in post-training. Reinforcement learning was built on the company's veRL framework, with a generative reward model used to judge the quality of answers.^[1]

The release also included a 2.4-billion-parameter Doubao ViT vision encoder. It has native dynamic-resolution image input. An end-to-end speech-to-speech framework was released alongside it.^[1]

A deep-thinking variant, Doubao-1.5-pro-AS1-Preview, was reported to surpass OpenAI o1-preview on AIME.^[1] AIME is a competition mathematics test that is often used to compare reasoning ability.

Price and serving

API prices for Doubao-1.5-pro-32k were 0.8 yuan per million input tokens, 0.16 yuan per million tokens for cache hits and 2 yuan per million output tokens. Doubao-1.5-lite-32k cost 0.3, 0.06 and 0.6 yuan under the same structure.^[2] Jiemian, citing an insider, reported that the model's sales through Volcano Engine carried a 50% gross margin.^[2]

Serving was optimised with prefill and decode separation on heterogeneous hardware, and with low-precision W4A8 quantization.^[1] Separating the two stages of generation lets each run on the hardware that suits it best, while low-precision weights reduce the memory and compute needed to serve each request.

Place in development

Doubao-1.5-pro came after the launch of the Doubao family, and its API was offered through Volcano Engine, the platform that sold the earlier Doubao models to enterprises.^[1][2] The most useful way to read its specifications is as a combination of three things: a sparse architecture that activates only part of the network for each token, a low API price, and input that spans text, images and speech.^[1]

The model's place in the family is therefore defined less by a single benchmark than by this mix of efficient architecture and cheap access. Readers comparing it with later models should keep in mind that the price figures above are per million tokens and were set at the time of its release.^[2]