Gemma 4 vs Qwen 3.6: Which Should You Run on Your GPU?

2026年7月5日

I dropped Google’s Gemma 4 and Alibaba’s Qwen 3.6 onto my home PC with an RTX 3090 and compared their performance. Both are open-weight LLMs (large language models) released in April 2026, and each comes in several sizes.

In this article I measured generation speed, output quality, math reasoning, and coding ability to figure out “which model fits your GPU’s VRAM." Bottom line: on a 24GB GPU, Qwen 3.6 27B was the best balance of quality and speed; on a 12GB GPU, Gemma 4 12B was.

Measured June 2026, via Ollama with Q4_K_M quantization.

Test setup

The models compared

In addition to every size of Gemma 4 and Qwen 3.6, I included the previous-generation Qwen 3 / 3.5.

Qwen 3.6 comes in only two sizes, 27B and 35B-A3B, with nothing in the 8–14B range. Since Qwen 3.6 isn’t an option on 12GB-or-less GPUs, you’ll be using the previous-generation Qwen 3 / 3.5 there.

Generation speed across all models

MoE (Mixture of Experts) models run only a fraction of their parameters per inference step, so they’re far faster than dense models. Gemma 4 26B MoE, with about 4B active parameters, hit 107 tok/s — beating the paid APIs GPT-5.5 (about 61 tok/s) and Claude Sonnet 4.6 (about 48 tok/s), a clear illustration of the speed advantage of local MoE models. That said, API speeds swing with server load and network conditions, so treat them as reference points only (speed data source: Artificial Analysis).

If you’re on a 24GB GPU (RTX 3090 / 4090)

With 24GB of VRAM, every model is on the table. The four main candidates were:

For quality, Qwen 3.6 27B. Its BenchLM overall score beat Gemma 4 31B, 73 to 64, and on coding benchmarks the gap was wide — an average of 70.6 vs 41.6. Speed is a practical 39 tok/s, and it fits on a single 24GB GPU. Its SWE-bench Verified pass rate was 77.2%. That still trails the paid Claude Opus 4.8 (88.6%) and GPT-5.5 (88.7%), but it’s a remarkable level for a model you can run locally for free.

For speed, Gemma 4 26B MoE. At 107 tok/s it’s about three times faster than Qwen 3.6 27B. Being MoE it has only 4B active parameters, but output quality and math reasoning had no practical issues. It’s plenty for chat and light writing.

For dense-model quality, Gemma 4 31B. This model runs all of its parameters rather than using MoE, and on benchmarks it came close to Qwen 3.6 27B. But at 24 tok/s it was the slowest, and even on a 24GB GPU it needs a 2-GPU split, so you have to be ready for the speed hit.

If you’re on a 12GB GPU (RTX 3060 / 4060)

With 12GB of VRAM, no size of Qwen 3.6 fits (even the smallest, 27B, needs about 19GB). Your options are Gemma 4’s smaller models or the previous-generation Qwen 3 / 3.5.

For quality, Gemma 4 12B. It’s the newest model in the 12B class, with the highest output quality and well-structured answers. Speed is a modest 35 tok/s, but it fit comfortably on a single 12GB GPU.

For speed, Gemma 4 E4B. It manages a practical 71 tok/s. It accepts image and audio input — both E4B and 12B of Gemma 4 handle audio (Qwen 3.5 9B supports image input). Output quality is slightly below Gemma 4 12B, but it was plenty for chat use.

If you’re on an 8GB-or-less GPU

On 8GB or less (RTX 3050, GTX 1070, and the like), the realistic options are Gemma 4 E4B (~6GB VRAM) or Qwen 3 8B (~6GB VRAM). Qwen 3 8B is the fastest at 127 tok/s, but as an older generation it lagged on benchmarks. If you don’t need multimodal, picking Qwen 3 8B for the speed is one reasonable move.

Output quality: a hands-on comparison

I compared with the same prompt: “the pros and cons of running a local LLM on a home PC, in about 500 words." Here are the answers from the two main 24GB-class models.

Gemma 4 31B’s answer

Qwen 3.6 27B’s answer

Gemma 4 31B tends to structure with bold text and headings, while Qwen 3.6 27B summarized the key points more concisely. Both produced natural prose, and both held up well on a formal-rewrite test too. Models of 12B and under are fine as far as basic language goes, but 27B-and-up models pull ahead in the depth of their answers and the richness of their examples.

Where they stand against paid models

How to read the benchmarks

A note on the benchmarks cited in this article.

Here’s how the models actually did on five sample GPQA Diamond questions (biology, chemistry, physics), run locally.

What stands out is that the MoE models (Gemma 4 26B MoE, Qwen 3.6 35B-A3B) matched the dense models at 80%. MoE has only 3–4B active parameters, but because it picks the best expert from all parameters for each token, its reasoning accuracy was higher than that small active count would suggest. Below 12B it fell to 40%, a clear sign of the model-size wall. The paid-model comparison below cites official benchmark results.

I’ve organized the gap between locally runnable models and paid cloud APIs using public benchmarks.

On GPQA Diamond (PhD-level science reasoning), the locally runnable Qwen 3.6 27B (87.8%) and Gemma 4 31B (85.7%) have closed in on Claude Sonnet 4.6 (89.9%). When I put Gemini 2.5 Flash through the same five-question sample via API, it scored 1/5 (20%). Flash is a speed-focused model, so it gives ground on reasoning accuracy. On SWE-bench Verified, Qwen 3.6 27B’s 77.2% trails Claude Opus 4.8 (88.6%) and GPT-5.5 (88.7%) — on coding, the paid models lead by more than 10 points.

Differences in Thinking mode

Both Gemma 4 31B and Qwen 3.6 27B have a Thinking mode (think before answering). In Ollama you control it with "think": true/false.

Their behavior differed. Gemma 4 31B generates about 1,000 characters of thinking before answering. Qwen 3.6 27B tended to generate more than 5,000 characters of detailed thinking. Thinking tokens count toward the generation cap (num_predict), so if you turn Thinking ON for Qwen 3.6 27B, you need to set num_predict to 2048 or higher — otherwise the thinking can eat the whole budget and the answer comes back empty.

Summary — the best model for each VRAM tier

If you have a 24GB GPU, Qwen 3.6 27B is the overall best. It’s well-balanced on speed, quality, and benchmarks, and on SWE-bench it showed performance approaching Claude Opus 4.8. Ideally, pair it with Gemma 4 26B MoE for the moments when you want speed.

On a 12GB GPU, Gemma 4 12B is the solid pick. Since Qwen 3.6 doesn’t offer a model in this VRAM tier, Gemma 4 is effectively the best option. When you need speed, switch to Gemma 4 E4B.

References

In 2026, open-weight models run at a level rivaling paid models if you have a 24GB GPU. Even at 12GB you can now get practical quality. Start by checking your GPU’s VRAM and picking a model from the tables above.

Hardware used for testing

【中古】MSI GeForce RTX 3090 GAMING X TRIO 24GB

¥148,000 (as of 2026/5/1)

Amazon楽天市場Yahoo!

【中古】ELSA GeForce RTX 3060 12GB

¥52,800 (as of 2026-06-22)

Amazon楽天市場Yahoo!