Ant Group open-sources Ling-3.0 base model weights
Two open-weight moves out of China show where the frontier labs think the value is moving — from finished chat models to the raw material underneath them, and from robot demos to robots that actually do shift work.
Ant Group's Bai Ling Lab has open-sourced the base weights of its Ling-3.0 models — and not just the polished, instruction-tuned chat versions. The release covers Ling-3.0-tiny-base and Ling-3.0-flash-base, plus six intermediate training checkpoints spanning pre-training, mid-training, and a WSM (warmup-stable-merge) stage that knits checkpoints together without a traditional learning-rate cooldown. The smaller model carries 7.9 billion total parameters with 1.3 billion active (a sparse MoE), while the larger flash variant runs 124 billion total / 5.1 billion active and is aimed at code, long-context, and agentic training. Both are live on Hugging Face and ModelScope.
What makes this more than a routine weight drop is the starting point it gives researchers. Most open releases hand you a model that's already been aligned and post-trained; Ant is instead shipping the clay — checkpoints you can continue pre-training, fine-tune, run preference optimization on, or use for RL rollouts from a known state. The lab frames it as a low-cost-to-large path: validate a training recipe on the 7.9B tiny base, then scale the method to the 124B flash base. The obvious caveat, which the company itself states, is that a base model is not a chat model — it isn't instruction-aligned and shouldn't be dropped into a user-facing product without further training and evaluation. For budget-constrained labs and RL research, though, that's exactly the asset that's been missing from most open-weight programs.
Galaxea used this week's World Robot Conference in Beijing to show embodied-AI robots doing unglamorous industrial work rather than parlor tricks. The star demos were a robot-staffed warehouse "front room" that takes online orders and autonomously picks, packs, and bags items — handling tens of thousands of SKUs and deformable plastic bags — and a robot assembling another robot, inserting centimeter-scale screws and driving them with a powered tool through a continuous vision-and-force-controlled sequence. Galaxea says reinforcement learning paired with an embodied foundation model lifted assembly speed by more than 30 percent over hand-programmed methods.
Alongside the demos, the company unveiled Nexo, a wheeled-arm humanoid with 30 degrees of freedom, 20 kg of arm payload, and roughly eight hours of battery life, and laid out a "1+3+N" strategy spanning its own hardware, an embodied base model, and industry deployments. The throughline is deliberately anti-spectacle: Galaxea's CEO argued the test of embodied AI isn't whether a robot can perform once on stage, but whether it recovers when a shelf is rearranged or a task is interrupted. That's the right bar — and the interesting question is whether show-floor competence turns into paid, boring, reliable shift work outside the expo hall.
What to watch: whether Ant's base-checkpoint release pushes other open-weight labs to ship training-stage weights by default, and whether Galaxea's demos convert into commercial deployments beyond the conference floor.
Should base-model checkpoints like Ant's become the default release for open-weight labs, or does handing researchers un-aligned clay create more risk than it's worth? Tell us in the comments.
Sources: Zhidx · Ling-3.0 base models (Hugging Face) · AI Era (Xin Zhi Yuan)
