Qwen 3.8 27B vs Gemma 4 31B

Analysis

Qwen 3.8 27B and Gemma 4 31B sit in the same weight class (27B and 30.7B parameters) and both solve the same problem, keeping a KV cache from growing without bound over a 262,144-token context, with genuinely different architectures. Qwen replaces most of its attention layers outright with linear attention; Gemma keeps every layer as real attention but caps most of them to a fixed window. Both strategies work, but they don't cost the same, and Qwen's own newly published benchmark card gives Gemma a real run of numbers to compare against for the first time.

Qwen 3.8 27B is dense with a hybrid stack: 16 of its 64 layers are Gated Attention (4 KV heads, 256 head dim) and keep a per-token KV cache; the other 48 run Gated DeltaNet, linear attention with a fixed-size recurrent state that costs the same at 2,000 tokens or 262,144. Gemma 4 31B is dense too, but keeps all 60 of its layers as real attention, capping 50 of them to a 1,024-token sliding window (google/gemma-4-31b-it's published config) while the remaining 10 full-attention layers use a narrower 2,048-wide KV entry (4 KV heads x 512 head dim) than the sliding-window layers' 4,096-wide one (16 x 256), an inversion carried over from Gemma 4 26B. At the full 262,144-token context both models natively support, Qwen's KV cache is 17.2 GB against Gemma's 22.3 GB — Qwen's more aggressive strategy (replacing attention layers rather than just windowing them) wins here despite Gemma's window being smaller than what Qwen's DeltaNet layers effectively achieve. Add each model's Q4_K_M weights and the full picture at that shared ceiling is 37.7 GB total for Qwen 3.8 27B against 45.9 GB for Gemma 4 31B — the difference between a 40 GB-class card and needing a 48 GB one. On benchmarks both report, Qwen leads: GPQA Diamond 89.2 vs Google's own self-reported 84.3, LiveCodeBench 90.3 vs 80.0. Qwen doesn't publish an MMLU-Pro score for the 27B release, so Gemma's 85.2 there has no direct comparison point on this site; Qwen's SWE-bench Pro (61.7) and Terminal-Bench 2.1 (73.0) are likewise numbers Gemma's own model card doesn't report, so those are Qwen's alone rather than a head-to-head. The two also differ in modality: Qwen 3.8 27B natively understands video as well as images; Gemma 4 31B is text-and-vision only. Both ship under an identical unrestricted Apache 2.0 license. Tooling currently favors Gemma: it has a working day-one Ollama tag (gemma4:31b) and four extra months in the wild, while Qwen 3.8 27B has no Ollama build yet at launch, making llama.cpp, vLLM, or SGLang the dependable route for it today.

Bottom line: On raw benchmark scores and on VRAM at long context, Qwen 3.8 27B is the stronger pick between these two: it needs about 8.2 GB less at the full 262,144-token window, scores higher on every benchmark both publish, and adds video understanding Gemma doesn't have. Gemma 4 31B's case is narrower but real: pick it if you need a working Ollama path today rather than waiting on Qwen's day-one tooling gap to close, if you specifically need the MMLU-Pro comparison point Qwen hasn't published for this release, or if you're already standardized on Google's Gemma ecosystem.