Kingy AI, a hardware-focused AI benchmarking outlet, ran three dense multimodal models, Qwen3.8-27B, Qwen3.6-27B and Gemma 4 31B-it, through identical workloads on a single RTX 4090 with 24GB of memory. The publication measured coding-agent completion, reasoning accuracy and, critically, how much VRAM survived as context length increased. Kingy AI’s commercial independence from Alibaba and Google is not established, so its numbers are best read as the publisher’s own measurements rather than independently verified benchmarks.
On matched Q4_K_M quantization with full GPU offload, Qwen3.8-27B completed all 12 coding tasks on the first seed and 35 of 36 seeded runs overall, against 16 of 36 for both Qwen3.6-27B and Gemma 4. It also passed 117 of 120 reasoning runs across three seeds and answered 23 of 24 document questions. None of that came at a speed cost: Qwen3.8 decoded at 49.09 tokens per second, essentially matching Qwen3.6’s 49.04.
The memory picture is where the three models split. At a 64,000-token context, Qwen3.8 and Qwen3.6 each peaked at 20,266 MiB with a standard F16 key-value cache, leaving 4.20 GiB free on the 24GB card. Gemma 4 hit a CUDA out-of-memory error at the same context length under the same cache setting. Kingy AI got it running only by switching Gemma’s cache to a lower-precision Q8_0 format, which pushed its peak to 21,956 MiB and left just 2.55 GiB of headroom.
Gemma 4 is not simply the weaker model. It posted a clean sweep on structured tool calling, 90 of 90 single calls and 30 of 30 multi-step runs, and edged out Qwen3.8 on a small vision test, 19 correct answers out of 20 against 18. Its iterative coding-agent runs were far less efficient, needing a median of ten tool calls and 44 seconds to reach a successful result, compared with five calls and 14 seconds for Qwen3.8-27B.
The buying decision this data actually supports is not “which model scores higher.” It is which model degrades gracefully as a session’s context window fills up. A team running short, schema-bound tool calls or image-heavy prompts can tolerate Gemma’s tighter memory margin. A team running a coding agent that needs to hold a growing conversation, edited files and test output in context cannot: Gemma’s need for a quantized cache at 64K is the kind of compromise that shows up as an unexpected crash once concurrency or a longer prompt enters production.
Kingy AI ran all three models on one physical GPU with a single llama.cpp build and a synthetic 50-task suite it designed itself, and it declined to build a composite score, arguing that weighting a coding task against a vision image would manufacture a preference rather than reveal one. That transparency is useful, but it does not substitute for reproduction on different hardware or a comparison against an established benchmark suite like SWE-bench.
Teams provisioning a single RTX 3090 or RTX 4090 for a local coding agent should default to Qwen3.8-27B, and reach for Gemma 4 only when the workload is dominated by declared tool schemas or image inputs. Anyone planning to push either model past 32K tokens of context should budget for a quantized key-value cache from the outset, rather than discovering the out-of-memory error after the system is already in production.
Based on reporting by Kingy AI (kingy.ai), published August 17, 2026.
