Model Bridging Lets a 753B LLM Think Silently While a 4B Model Answers: Inference Costs Cut to One-Twentieth

星战科技团队

2026-09-10 09:10:00

48 次观看

After GPT-6 Astra scored near-perfect on ARC-AGI 3, Mostik—a four-month-old team led by Fields Medalist Stanislav Smirnov—used “model bridging” to deliver striking results: cloud-based 753B GLM-5.2 handles only prefill and emits no text, while an on-device 4B Qwen-3.5 produces the entire output. The approach reportedly closes about 50% of the performance gap and cuts large-model-side inference costs to roughly one-twentieth, pointing to reduced text handoffs as a new cost-reduction path for...

After GPT-6 Astra posted near-perfect results on ARC-AGI 3, a 4B small model that can run on a phone also delivered a remarkable score on the same benchmark.

trictly speaking, this is a combination of 4B Qwen-3.5 and 753B GLM-5.2: GLM-5.2 stays in the cloud, Qwen-3.5 runs on the phone, the large model never outputs a single word, and the