Meta shipped Muse Glimmer under Apache 2.0 on August 10, 2026. Roughly 24 hours later, NVIDIA released Nemotron 3.5 Lightning under the OpenMDW-1.1 license. Both are 30B-class open-weight models aimed squarely at the same target: agent execution running locally on consumer GPUs, not orchestrated through a closed cloud API.

The design choices differ, and the differences are the story.

Muse Glimmer is a dense 30B model that Meta Superintelligence Labs distilled from its proprietary Muse Spark via logit distillation. Full precision lands around 55GB. Four-bit quantization brings it under 20GB with what SiliconANGLE describes as "minimal to no degradation on agentic tasks," alongside integrations for llama.cpp, MLX, and ExecuTorch. Meta evaluated it across roughly two dozen benchmarks and reported outperforming Gemma4-31B and Qwen3.6-27B on about half of them. Artificial Analysis clocks Glimmer at 35 on its Intelligence Index.

Nemotron 3.5 Lightning takes the opposite architectural bet. It's a hybrid Mamba MoE inheriting the shape of Nemotron 3 Nano 30B A3B: 31.6B total parameters, 3.6B active. On a pre-release DeepInfra endpoint running NVFP4 weights, Artificial Analysis measured a median output speed of nearly 670 tokens per second. The Intelligence Index score of 24 is a +9 jump over Nano and sits just two points below Nemotron 3 Super, a model roughly four times its size. On Terminal-Bench v2.1 it moves from Nano's 7% to 24%. GDPval-AA v2 Elo climbs 334 points to 824. At about one quarter the size of gpt-oss-120b, it matches that model on the Intelligence Index.

NVIDIA's framing in its developer blog is telling. Lightning is positioned as the small member of a routable family, wired through NeMo Switchyard, where frontier models handle orchestration and Lightning absorbs the high-volume tool calls. That's not a chatbot pitch. It's an execution-layer pitch, and it assumes the agent stack itself is decomposing into specialized tiers.

That this is NVIDIA's first open-source release since Jensen Huang publicly backed the approach in late July isn't incidental, and neither is the timing against Meta. The day before Lightning shipped, Mark Zuckerberg published an essay arguing American open-source models "should be the best globally." The two companies are now openly competing to define what that best looks like at the 30B tier, one via dense distillation from a closed teacher, the other via sparse routing built for throughput.

Glimmer is smarter on paper. Lightning is faster and cheaper to run per token. The execution layer, the part of the agent stack that actually calls the tools, is being priced and shipped as commodity infrastructure by the two American labs with the most to lose if it isn't.

Meta and NVIDIA drop competing 30B open-weight agentic models within 24 hours

Sources