GLM-5, MiniMax M2.5 and Seedance 2.0 arrive before Lunar New Year

Chinese labs tend to ship before the Lunar New Year break, and they did.

Seedance 2.0

The thing everyone was sharing was Seedance 2.0 from ByteDance. It is a video model that accepts several kinds of reference at once, text, images, audio clips and video, and mixes them into one generation. The clips that circulated had steadier motion and better scene-to-scene continuity than most of what I had seen from earlier models.

For anyone making video for a living, the interesting part is control. Early text-to-video was a slot machine. Giving the model references to follow is how you get a result you meant. It also makes misuse easier, and ByteDance faced immediate questions about likeness and copyright. I would expect tighter rules from the platform soon.

Open models

Zhipu's GLM-5 is the bigger of the two open flagships, positioned around agentic engineering and long tasks. MiniMax M2.5 came out a day later with an emphasis on running agents cheaply for long stretches. Both are the sort of model that makes me rethink which parts of my agent stack really need a closed API.

Frontier updates

Google updated Gemini 3 Deep Think, its slower reasoning mode, with a focus on research-grade science and engineering problems. OpenAI released GPT-5.3-Codex-Spark, a fast Codex variant running on Cerebras hardware, built for quick interactive edits instead of long runs.

Money

Anthropic announced a $30 billion Series G at a $380 billion post-money valuation. The size of it is the signal: the labs expect inference demand from agents to keep growing and are raising to meet it.

Looking at the open models, I want to flag how fast the Chinese labs are iterating. GLM went from 4.7 to 5 in under two months, and MiniMax moved from M2.1 to M2.5 in about the same time. For a buyer, that pace means evaluating on a schedule, not on a hunch. I put a reminder in my calendar to retest on the first of each month, and the results go into one table I can show a client.

What I'd watch next: what the video platforms do about reference-based generation and likeness, and how quickly GLM-5 shows up in third-party coding tools.