ByteDance's Doubao Large Model 2.2 Release Postponed from August

IT Home, August 30 — ByteDance's Doubao large model 2.2, originally scheduled for release in August, will be postponed, according to Whale Lab. Multiple sources close to ByteDance stated that the reason for the delay is the company's internal desire to enhance coding, tool-calling, and agent capabilities through more comprehensive pre-training and post-training. The goal is to trade a longer R&D cycle for a more significant capability improvement. Seed is currently facing two simultaneous tasks: advancing the next-generation ultra-large model while continuing the normal iteration of existing models. Therefore, version 2.2 is more like a crucial reinforcement sandwiched between two generations of models. It cannot wait for the next-generation ultra-large model to be completed before addressing coding, nor can it simply undergo another basic iteration using past methods. Elevating its coding capabilities to the first tier is one of the Seed team's primary goals for 2026. Reports indicate ByteDance hopes the Seed model's coding capabilities will generate an industry impact similar to Zhipu's GLM-5.2 and Kimi-K3 before the end of the year. According to a report by The Information on August 5, ByteDance founder Zhang Yiming explicitly stated at last month's Seed team all-hands meeting that the company will not rely on AI distillation technology to improve its models, even if it means temporarily falling behind competitors. IT Home noted that on August 6, ByteDance CEO Liang Rubo stated at the 2026 mid-year all-hands meeting that for large language models, ByteDance will continue to insist on in-house R&D, focus on the basics, accept short-term lag, and persist in long-term optimization. Related readings: "Paving the way for a 5-trillion-parameter ultra-large model? Reports say ByteDance adjusts Seed foundation model team structure" "Largest known scale in China! Reports say ByteDance is discussing training a model with over 5 trillion parameters" "ByteDance Doubao Seed 2.1 Pro and Turbo deep thinking models released, three major capabilities on par with GPT-5.5" "ByteDance Doubao large model 2.0 released: Expert mode launched, most benchmarks reach SOTA levels"