New open-weight models for 32-64GB VRAM: Google Gemma 4, Tencent Hy3, and Muse Spark 1.1; SAGE benchmark highlights
Key Questions
What are the key features of Google's Gemma 4 model?
Gemma 4 is released in sizes up to 31B parameters that fit in 32-64GB VRAM and rivals much larger models. On the SAGE benchmark it leads open-source models with a 55.03% score, nearly matching Claude Opus 4.7.
How does Tencent's Hy3 model compare in size and licensing?
Hy3 is a 295B MoE model with 21B active parameters released under Apache 2.0. It offers cost-effective performance suitable for local deployment.
What endorsements has Muse Spark 1.1 received for agentic use cases?
It is praised for strong performance in harnesses like OpenClaw and HermesAgent, with enterprise validation from Box evaluations. Users note it can outperform models like Opus 4.8 at roughly 20% of the cost.
Google releases Gemma 4 technical report (up to 31B, fits VRAM, rivals larger models). SAGE benchmark: Gemma 4 31b leads open-source at 55.03%, nearly matching Claude Opus 4.7 (56.10%). Tencent releases Hy3 (295B MoE, 21B active, Apache 2.0, cheap pricing). Muse Spark 1.1 shows RL scaling improvements and gets endorsements for agent harnesses (OpenClaw, HermesAgent) and enterprise praise from Box evaluation, but open-source status and VRAM requirements still unclear. All are strong contenders for local deployment and agentic coding.