ByteDance's Doubao Large Model Unveiled with Multiple Versions and Pricing Details

TapTechNews May 21st news, ByteDance launched the Doubao large model (originally named Yunque) at the Spring Volcano Engine FORCE Prime Mover Conference on May 15th. This model is mainly for industry scenarios and can provide a variety of graphic, audio and text generation capabilities. TapTechNews sorts them out as follows:

Doubao General Model pro: ByteDance's self-developed LLM model professional version, supports 128k long text, the whole series can be precisely adjusted, and has stronger comprehensive abilities such as understanding, generation, and logic, and is suitable for rich scenarios such as question answering, summarizing, creating, and classifying;

Doubao General Model lite: ByteDance's self-developed LLM model lightweight version, compared with the professional version, it provides lower token cost and lower latency, and provides a flexible and economical model choice for enterprises;

Doubao · Role-playing Model: Personalized role creation ability, stronger context awareness and plot promotion ability to meet the flexible role-playing needs;

Doubao · Speech Synthesis Model: It provides natural and vivid speech synthesis ability, is good at expressing various emotions and deducing various scenes;

Doubao · Sound Replica Model: It can achieve 1:1 cloning of sound within 5 seconds, highly restores the similarity and naturalness of timbre, and supports cross-language migration of sound;

Doubao · Speech Recognition Model: Higher accuracy and sensitivity, lower speech recognition latency, and supports correct recognition of multiple languages;

Doubao · Image-generation Model: More accurate text understanding ability, more accurate graphic matching, and more beautiful picture effect, and is good at creating elements related to Chinese culture;

Doubao · Functioncall Model: It provides more accurate function recognition and parameter extraction ability, and is suitable for complex tool invocation scenarios;

Doubao · Vectorization Model: Focuses on the usage scenario of vector retrieval, provides core understanding ability for the LLM knowledge base, and supports multiple languages.

Today, the official website of the Volcano Engine updated the pricing details of the Doubao large model, claiming that on the basis that the model inference pricing is significantly lower than the industry price, the TPM and RPM of the Doubao General Model have reached the highest domestic standards, and the price is 99% lower than the industry, and the TPM limit is 2.7 to 8 times that of the same-specification model; In addition, the relevant models can also use the prepaid and postpaid models:

Taking the Doubao General Model pro-32k as an example: According to the unit price calculation of the prepaid model, the monthly price of 10KTPM is 2000 yuan. 10K*60*24*30=43200K.

That is, the price of 432000KTokens is 2000 yuan, and the average price is 0.0046 yuan/thousand Tokens. According to the calculation of 0.0008 yuan/thousand Tokens for the inference input and 0.002 yuan/thousand Tokens for the inference output of the Doubao General Model pro-32k, the comprehensive price of the model inference is 0.001 yuan/thousand Tokens.

The official said that the TPM limit of other domestic competing models is mos tly between 100K and 300K, and the RPM is in the range of 60 to 120. The RPM limit of the lightweight model is relatively high, but only between 300 and 500. According to the 10KRPM limit calculation, enterprise customers can simultaneously call the Doubao General Model 167 times per second on average, so as to meet the large model application needs in the production system in most business scenarios.

The official also emphasized that the relevant standards have reached the RPM upper limit provided by OpenAI for high-level customers (Tier4 and Tier5 level customers). On the long-text model with greater computing power challenges, the 128k versions of the Doubao General Model pro and lite have a model current limit of 1KRPM and 400KTPM, which is also significantly higher than that of other domestic 128k long-text models, which can help enterprises use large models at a lower cost and accelerate the application of large models.