Nemotron 3.5 Lightning 30B-A3B vs Muse Glimmer 30B

Analysis

Nemotron 3.5 Lightning and Muse Glimmer 30B launched one day apart in August 2026 (Nemotron on the 11th, Muse Glimmer on the 10th) and land at almost the same total parameter count, 30B against 27.8B, which makes how each spends that budget the real story. Nemotron is a sparse MoE model that activates only 3B parameters per token by routing through experts and skipping attention on most layers entirely. Muse Glimmer is fully dense: every one of its 27.8B parameters fires on every token. That gap in active compute doesn't translate the way it might seem to once VRAM, rather than speed, is the question.

Nemotron routes through 128 experts (6 active plus 1 shared per token, 3B of 30B parameters active), and only 6 of its 52 layers keep a KV cache at all (2 KV heads x 128 head dim, 256-wide); the rest split between 26 Mamba-2 layers and 20 MoE FFN layers with nothing to cache. Muse Glimmer is dense end to end, all 27.8B parameters active on every token, and all 52 of its layers are real attention, but 39 of them (three sliding-window layers per one full-attention layer) cap at a 2,048-token window, with every layer sharing the same narrow 2 KV heads x 128 head dim, 256-wide entry, one of the tightest GQA ratios (16:1) tracked on this site. At the 131,072-token context where Muse Glimmer's own window ends (Nemotron's continues on to 1,048,576), Nemotron's KV cache is 0.8 GB against Muse Glimmer's 1.8 GB, still the smaller of the two, since it caches only 6 of 52 layers against Muse Glimmer's 13 of 52 uncapped layers at the same per-layer width. But add each model's Q4_K_M weights and the ranking flips: 21.4 GB total for Nemotron against 21.0 GB for Muse Glimmer. Nemotron's Q4_K_M weights alone are 18.3 GB against Muse Glimmer's 16.9 GB, because weight size tracks total parameters, not active ones, Nemotron's MoE experts all have to sit resident in VRAM even though only a handful fire per token, so being 3B-active buys inference speed, not a smaller download. The same ordering holds at a shorter 8,192-token context: 20.5 GB for Nemotron against 19.2 GB for Muse Glimmer. On quality, Muse Glimmer leads every benchmark the two share, and by a wide margin: GPQA Diamond 83.5 vs 75.57, SWE-bench Verified 76.0 vs 52.80, and Terminal-Bench 2.1 51.7 vs 23.46, more than double Nemotron's score on that last one. That gap tracks the active-parameter gap directly: Muse Glimmer computes with all 27.8B of its parameters on every token, Nemotron with roughly a tenth that. NVIDIA isn't positioning Lightning to win on quality; its own figures claim up to 4x higher throughput and 30% faster task completion against similarly-sized open models, a tradeoff Muse Glimmer's fully dense design doesn't make in the first place. Licensing is Apache-family on both sides (Nemotron under NVIDIA's own OpenMDW-1.1, Muse Glimmer under standard Apache 2.0), and Muse Glimmer additionally accepts image input via a separate encoder where Nemotron is text-only. Neither model had a clean day-one tooling story: Muse Glimmer's only launch-day Ollama tag runs exclusively through Apple's MLX engine, with NVIDIA and AMD support still rolling out at release, while Nemotron shipped without any GGUF or Ollama build at all, native NVFP4, FP8, and BF16 checkpoints only. llama.cpp or vLLM is the dependable route for both on non-Apple hardware, once a GGUF exists for Nemotron in the first place.

Bottom line: Muse Glimmer 30B is the stronger pick on quality alone, leading every shared benchmark by a wide margin, and once weights are counted it's also the smaller download, 16.9 GB against Nemotron's 18.3 GB at Q4_K_M. Nemotron 3.5 Lightning's case is narrower: pick it for raw agentic throughput at scale, NVIDIA's own 4x claim is the entire pitch, or if a workload genuinely needs a context window past Muse Glimmer's 131,072-token ceiling, since Nemotron's continues on to 1,048,576. Within the context range these two models share, both fit comfortably on a single 24 GB card.