[X]score: 0.76

25+ Open-Weight Model Drops: Nemotron 550B, Gemma 4 12B, and More

June 5, 2026

A single week saw major open-weight releases including NVIDIA Nemotron Ultra at 550B parameters (55B active, MMLU 89.1, 1M context), Google Gemma 4 12B (256k context, multimodal, AIME 2026 77.5), and Liquid AI LFM2.5-8B with only 1.5B active parameters hitting MATH500 88.8. Sparse MoE architectures dominate, with multiple Apache 2.0 releases closing the gap on closed frontier models.

HOW THIS AFFECTS YOU

●

builderMultiple production-ready open-weight options are now available across modalities — Gemma 4 12B with QAT/ONNX/MLX packaging and Liquid LFM2.5 for on-device deployment are the most immediately shippable.

●

researcherNemotron Ultra's hybrid Mamba-MoE at 550B with only 55B active parameters and 1M context is the most architecturally novel release, worth benchmarking against closed models.

●

founderThe volume and quality of open-weight releases this week materially reduces dependency on closed API providers for most product use cases.

●

investorOpen-weight model quality is converging with closed frontier models faster than expected, compressing the moat of API-only providers.

read original ↗x.com

← back to feed