Nemotron 3.5 Lightning 30B-A3B vs Qwen 3.6 27B

Analysis

Qwen 3.6 27B and Nemotron 3.5 Lightning 30B-A3B are two agent-oriented releases four months apart in 2026 (Qwen in April, Nemotron in August) that both chase a small KV cache and a context window well past 100,000 tokens, but by very different means. Qwen is a dense model that replaces most of its attention layers with linear-attention DeltaNet blocks; Nemotron goes further still, mixing Mamba-2 state-space layers, MoE feed-forward layers, and a handful of ordinary attention layers in the same 52-layer stack. The more consequential difference is how much of each model is actually doing work per token: all 27B of Qwen's parameters fire on every token, while only 3B of Nemotron's 30B do.

Qwen 3.6 27B is dense: 16 of its 64 layers keep a real KV cache (4 KV heads x 256 head dim, 1,024-wide), and the other 48 run Gated DeltaNet, a fixed-size recurrent state instead of attention. Nemotron 3.5 Lightning is sparser on both axes at once: it's MoE (128 routed experts, 6 active plus 1 shared per token, so only 3B of its 30B parameters are active on any given token), and only 6 of its 52 layers keep a cache at all (2 KV heads x 128 head dim, 256-wide, a quarter of Qwen's per-layer width), the rest split between 26 Mamba-2 layers and 20 MoE FFN layers with nothing to cache. At the 262,144-token context where Qwen's native window ends (Nemotron's continues on to 1,048,576), Nemotron's KV cache is 1.6 GB against Qwen's 17.2 GB, an order of magnitude smaller, because Nemotron caches roughly a tenth of its layers (6 of 52) against Qwen's quarter (16 of 64), at a narrower width besides. Add each model's Q4_K_M weights and the full picture is 22.3 GB total for Nemotron against 37.7 GB for Qwen, the difference between a 24 GB card with room to spare and needing 32 GB or more. On benchmarks both report, Qwen leads clearly: MMLU-Pro 86.2 vs 81.62, GPQA Diamond 87.8 vs 75.57, SWE-bench Verified 77.2 vs 52.80. Nemotron's own reported Terminal-Bench score, 23.46, is measured on the 2.1 suite; Qwen's reported 59.3 is on the older 2.0 suite, a different harness, so the two aren't a direct head-to-head. NVIDIA isn't positioning Lightning as a quality leader in the first place: the tradeoff for routing through only 3B active parameters is throughput, and NVIDIA's own figures claim up to 4x higher output and 30% faster task completion against similarly-sized open models. Licensing differs too: Nemotron ships under NVIDIA's own OpenMDW-1.1, Qwen under the more widely recognized Apache 2.0. Qwen also accepts image and video input where Nemotron is text-only, though Nemotron's native context window (1,048,576 tokens) is four times Qwen's (262,144), which itself only reaches roughly 1,010,000 with YaRN, a context-stretching technique applied after training. Tooling favors Qwen by default: it shipped four months earlier, giving its llama.cpp and vLLM support time to mature even though its own day-one GGUFs reportedly didn't load in Ollama either; Nemotron launched on 11 August 2026 with no GGUF or Ollama build at all, native NVFP4, FP8, and BF16 checkpoints for vLLM, SGLang, and TensorRT-LLM only.

Bottom line: For long-context or throughput-bound agent workloads on a single 24 GB card, Nemotron 3.5 Lightning is the more VRAM-efficient design, and NVIDIA's own throughput claims are the reason it trades benchmark quality away in the first place. Qwen 3.6 27B is the better default otherwise: it scores higher on every shared benchmark, adds vision and video input Nemotron doesn't have, ships under a fully permissive and widely recognized license, and already has four months of runtime support behind it that Nemotron's day-one release doesn't.