Baidu has open-sourced ERNIE 5.5 Flash Base, a 280-billion-parameter mixture-of-experts model pre-trained on 33.5 trillion tokens, according to The Decoder. The notable part is not the size. It is what Baidu left out.
The model ships before any fine-tuning. MarkTechPost reports it was published ahead of mid- or post-training, with 14 billion active parameters at inference. The Decoder frames this as deliberate: Baidu wants developers to take the base model wherever they want, rather than shipping a product with Baidu's choices already built in.
A week of raw building blocks
ERNIE 5.5 Flash Base is not alone. Thinking Machines Lab released Inkling 2 Small, a 594-billion-parameter mixture-of-experts model with 17 billion active parameters, with Apache 2.0 weights planned, per MarkTechPost. Tencent released Hy4, described by The Decoder as its first multimodal model able to understand images and operate software on its own, a capability known as computer use.
Three different labs, three different architectures, one shared move: ship a large, general-purpose model as infrastructure rather than as a finished assistant.
For teams building on open humanoid robot platforms, that distinction matters. A base model with no fine-tuning is not a capability a robot has today. It is a starting point: grounding it in sensor data, teaching it to issue motor commands, and testing it against failure modes that only show up on real hardware is work a robotics team still has to do itself. None of this week's releases is a robot control system.
Why the gap between model and autonomy is still wide
The clearest evidence for that gap this week comes from outside robotics entirely. SemiAnalysis launched GDPx, a benchmark designed to measure how AI agents perform on real-world work tasks rather than curated test sets. Per The Decoder, no AI agent scored above 12 percent. Claude Opus 5.5 led the field, narrowly ahead of Gemini.
That is a benchmark for general software agents, not humanoid robots. But it is relevant to anyone evaluating robot platforms on the strength of the language or vision models underneath them. If the most capable agents available today still score in single digits on a controlled, real-world benchmark, a robot's marketing claim of being "built on" a large new foundation model says little about what it can reliably do in a warehouse, a home or a factory. The model is necessary. It is nowhere near sufficient.
What this means for comparing platforms
Treat "built on a large open model" as a roadmap signal, not a capability claim. It tells you a platform has access to serious raw material and the flexibility to adapt it. It does not tell you whether that platform's autonomy stack can pick up an unfamiliar object, recover from a dropped grip, or follow a multi-step instruction in a cluttered room.
Those questions separate working products from demos, and they are answered by task-level evidence, not parameter counts. Until a robotics team publishes real task success rates built on these newly released base models, ERNIE 5.5 Flash Base, Inkling 2 Small and Hy4 are inputs to a future generation of humanoid software, not evidence that one has already arrived.
Readers tracking where that evidence shows up can follow the humanoid robot price tracker and ongoing robotics and AI coverage.
Sources: The Decoder, MarkTechPost
Written with the help of AI tools, from published sources, under our editorial rules. Illustration generated with AI. No commission on purchases. Learn more


