Sage-A1.5 MoE Agent Hits 81.7% BrowseComp, 61.2% SWE-Bench Pro

Sage-A1.5 is a 35-billion-parameter mixture-of-experts agent with only 3B parameters active per token. On BrowseComp it scores 81.7, against Agents-A1's 75.5; on SWE-Bench Pro the jump is sharper, from 43.6 to 61.2. On GAIA it lands just under its predecessor at 95.1 to 96.0.

The two additions over Agents-A1, per the paper on Hugging Face, are an orchestration layer that coordinates subagents and a verification protocol toggled by a 'Verification protocol:enabled' marker in the system prompt. Enabling it on BrowseComp takes accuracy from 81.0 to 85.5 while cutting output tokens by 5.52% and turns by 3.15%. 'The central idea is to make verification a learned capability of the model itself,' the authors write, so that 'the model adaptively determines when to verify, what to verify, and how to continue based on the findings, rather than having verification triggered and executed by an external harness according to predefined rules.'

FP8 and 4-bit quantization bring the active footprint down to 3B, with runnable targets listed from NVIDIA A100 through B200, AMD MI300X, Huawei Ascend 910B/910C, and Apple M-series, via LM Studio, llama.cpp, or Apple MLX.

The claim doesn't carry everywhere. On BrowseComp outright, Kimi-K3 (91.2) and GPT-5.6-sol (90.4) still beat it, and on HLE-with-tools and CritPt the larger frontier models lead. This is the kind of release our agents tracker has been clocking at near-weekly pace, mostly on orchestration-and-verification architectures.