Feb 12, 2026 · 24-Hour AI Briefing: ByteDance’s Seedance 2.0 Goes Global, Doubao 2.0 Set for Feb 14; Zhipu’s GLM-5 (744B MoE) Goes Fully Open Under MIT
In the last 24 hours, two storylines accelerated in parallel: ByteDance is pushing AI video into a more “system-level” capability (native audio-video sync + multi-shot narrative control), and Zhipu is betting on a massive MoE model that keeps inference cost manageable—then removing friction entirely by releasing weights under the MIT License.
1. ByteDance draws global attention with Seedance 2.0 and announces Doubao Large Model 2.0 for Feb 14
Commentary:
Seedance 2.0’s native audio-video synchronization, multi-shot storytelling, and voiceprint-style human voice replication are being framed by many as a step-change in AI video. Native A/V generation directly attacks the “audio-video mismatch” problem, while model-driven interpolation of intermediate motion helps in tightly controlled use cases (ad-style stop-motion, scene continuity, story stitching). These are exactly the capabilities that typically widen the gap in physics plausibility, motion coherence, and cross-shot consistency.
With Seedance 2.0 pulling attention, ByteDance is now scheduling Doubao 2.0. Doubao reportedly already exceeds 100M DAU in China, which matters: it’s not just a model, it’s a product-scale distribution surface with real feedback loops. If Doubao 2.0 meaningfully upgrades enterprise-grade agent capabilities and multimodal input support, the value proposition shifts quickly from “chat” to “deliverable workflow productivity.”
2. Zhipu releases GLM-5: 744B total parameters (≈40–44B active), optimized for coding and agent tasks, and fully open-sources weights under the MIT License
Commentary:
GLM-5 uses a Mixture-of-Experts (MoE) architecture: huge total capacity, but only ~40–44B activated per inference step. That’s the “big but efficient” play—push model capacity up while keeping real compute cost closer to a mid-sized dense model, which is friendlier for deployment and concurrency.
Its strengths are positioned around coding and agent workloads, where tool-use reliability, workflow orchestration, error recovery, and long-context stability matter more than generic chat polish. General conversation and broad multimodal understanding may still lag top-tier baselines like GPT-5 or Claude Opus 4.6.
The mention of a Slime RL framework, async agent RL methods, and sparse attention mechanics points to an engineering focus on learning efficiency and runtime support. The headline move is licensing: releasing weights under MIT dramatically lowers adoption friction for both research and commercial deployment, making ecosystem expansion far easier.
Most important AI events in the last 72 hours
The pattern is getting clearer: one race is about product-scale distribution and closed-loop iteration, the other is about open licensing and enterprise deployability. The next concrete signal will be what Doubao 2.0 actually ships for agents—and how fast GLM-5 spreads in real production environments under MIT.