Qwen 3.8 27B vs Nemotron 3.5 Lightning 30B-A3B

Analysis

Qwen 3.8 27B and Nemotron 3.5 Lightning 30B-A3B are both August 2026 agentic releases, Nemotron on the 11th, Qwen on the 14th, built around the same goal of keeping a KV cache small enough for a single consumer GPU, reached through opposite means. Qwen is dense, with most of its attention layers swapped for linear attention; Nemotron is sparse MoE, with most of its layers swapped for Mamba-2 state-space blocks that skip attention entirely. The more consequential difference is how much of each model actually computes per token: all 27B of Qwen's parameters fire on every token, only 3B of Nemotron's 30B do.

Qwen 3.8 27B keeps 16 of its 64 layers as Gated Attention (4 KV heads, 256 head dim), the only layers that accumulate a KV cache; the other 48 run Gated DeltaNet, linear attention with a fixed-size recurrent state. Nemotron 3.5 Lightning is sparser on two 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 token, and only 6 of its 52 layers keep a cache at all (2 KV heads, 128 head dim, a quarter of Qwen's per-layer width), the rest split between 26 Mamba-2 layers and 20 MoE feed-forward layers with nothing to cache. Qwen's native context window ends at 262,144 tokens, a quarter of Nemotron's 1,048,576; restricted to the range both can reach, Qwen's KV cache is 17.2 GB against Nemotron's 1.6 GB, an order of magnitude smaller for Nemotron even though Qwen already replaced most of its own layers with linear attention, because Nemotron caches roughly a ninth of its layers (6 of 52) at a quarter of Qwen's per-layer width. Add each model's Q4_K_M weights and the full picture is 37.7 GB total for Qwen 3.8 27B against 30.3 GB for Nemotron 3.5 Lightning, the difference between needing a 40 GB-class card and just fitting a 32 GB one. Nemotron's real Q4_K_M file is 25.5 GB, so it does not fit a 24 GB card at this context either. On the two benchmark fields both models report, Qwen leads by a wide margin: GPQA Diamond 89.2 against 75.57, and Terminal-Bench 2.1 73.0 against 23.46, both explicitly the 2.1 suite on each vendor's own model card, so the comparison holds even though the two scores come from different labs' own harnesses. NVIDIA isn't chasing benchmark leadership with Nemotron's 3B-active design in the first place; the pitch is throughput, and NVIDIA's own figures claim up to 4x higher output and 30% faster task completion against similarly-sized open models, the tradeoff that routing through a fraction of the parameter count buys. Licensing differs too: Nemotron ships under NVIDIA's own OpenMDW-1.1, Qwen under the more widely recognized Apache 2.0 with no usage restrictions. Modality splits as well: Qwen 3.8 27B accepts image and video input where Nemotron is text-only. Tooling favored neither model cleanly at launch: Qwen 3.8 27B shipped without an Ollama tag, and Nemotron with only native NVFP4, FP8, and BF16 checkpoints for vLLM, SGLang, and TensorRT-LLM. Both have since gained Ollama tags (qwen3.8:27b and nemotron-3.5-lightning:30b) and Nemotron has community GGUFs, so llama.cpp and Ollama are now options for either.

Bottom line: For quality, multimodal input, or a workload that stays under 262,144 tokens, Qwen 3.8 27B is the stronger default: it leads both shared benchmarks by a wide margin and needs no exotic runtime to try. Nemotron 3.5 Lightning's case is narrower but real: pick it for throughput-bound agent workloads on a single 32 GB card, NVIDIA's own 4x claim is the entire pitch, or for a context window Qwen simply can't reach, Nemotron's native 1,048,576 tokens is four times Qwen's ceiling.