Gemma 4 or Qwen: which local model writes better English?
The two open-weight families most people download, run on our own machine and judged on the same English prompts in blind duels: which one writes better, in which kinds of text, and what it costs in memory.
The verdict
Qwen3.8 27B by Alibaba (Q4_K_M, thinking xhigh) writes better English than Gemma 4 31B by Google (Q6_K, thinking on) in our blind duels: rank 40 against rank 46 on the leaderboard, score 58.7 against 53.9.
Head to head, Gemma 4 takes 35.4% of the points over 24 direct duels (7 wins, 3 ties, 14 losses). Gemma 4 leads in 4 of the 10 kinds of writing where both have been judged, Qwen in 6.
| Model | Rank | Score | Duels |
|---|---|---|---|
Gemma 4  Gemma 4 31B 💡 🏠 | 46 | 53.9 | 2,289 |
Qwen  Qwen3.8 27B 💡 🏠 | 40 | 58.7 | 2,279 |
Each side is its best-ranked variant in this edition: Gemma 4 31B by Google (Q6_K, thinking on) for Gemma 4, Qwen3.8 27B by Alibaba (Q4_K_M, thinking xhigh) for Qwen. When a better variant of either arrives, this page follows it.
Kind of writing by kind of writing
Score and rank recomputed on the duels of each category alone; the leader is in colour. Click a kind of writing for its full ranking.
| Kind of writing | Gemma 4 | Qwen | Direct duels |
|---|---|---|---|
| Articles, blogs and editorial | 33.8 #75 | 89.1 #10 | — |
| Dialogue and screenplay | 87.5 #19 | 26.0 #73 | — |
| Explanation and popularization | 40.2 #71 | 64.7 #33 | 1–0–0 |
| Fiction and narrative | 63.7 #37 | 92.9 #14 | — |
| Marketing and sales | 56.6 #46 | 27.1 #80 | — |
| Personal communication | 51.4 #48 | 56.6 #43 | 5–3–12 |
| Professional and administrative | 62.5 #43 | 65.1 #40 | 0–0–1 |
| Social media and short formats | 61.9 #38 | 82.1 #20 | 0–0–1 |
| Speech and public speaking | 63.5 #39 | 41.1 #55 | 1–0–0 |
| Structured documents and reports | 86.0 #11 | 70.8 #34 | — |
Direct duels read wins–ties–losses for the first model. A dash means the model has not yet been judged in that category.
What each side stands for
“Gemma 4” is Google’s open-weight family, “Qwen” Alibaba’s, and each side is represented by its best-ranked file in this edition: a given size and a given quantization, run in LM Studio on our Framework Desktop with reasoning on where the file supports it. The table above names the two files. The other sizes and quantizations of both families, with what each needs in memory, are on the Local AI page.
What the numbers measure
Blind duels: two anonymous texts for the same English prompt, an AI judge picking the better one without seeing the names. Local files meet cloud models in the same pool, so their scores are on the same scale as the rest of the leaderboard. The direct record counts only the duels where these two files met.
How to choose
For a local model the decision is rarely quality alone. Qwen’s best files need a reasoning pass that makes them slow; Gemma 4 comes in sizes from 2 to 31 billion parameters and its mixture-of-experts build writes fast. If the two are close on the kind of writing you do, take the file that fits your memory with room for context, then check the quantization and reasoning articles: for these two families both settings move the score by tens of points.
Writing example
Prompt
Write a very short toast (50–150 words) to be delivered by the best man, Sam, at the wedding reception of his childhood friends Priya and Leo. Include a light, affectionate memory of the three of them building a leaky raft one summer, a warm compliment about how well Priya and Leo balance each other, and a heartfelt wish for their future. End by inviting the guests to raise their glasses. Keep the tone celebratory, sincere, and suitable to say aloud.
Published verdict: Gemma 4 (judged by Claude Opus 5.5).
Gemma 4 · Gemma 4 31B 💡 🏠
Hi everyone, I’m Sam. Growing up with Priya and Leo was always an adventure—most notably the summer we decided to build our own raft. It leaked almost immediately, leaving us soaked and laughing as we desperately tried to stay afloat.
That same spirit of joy is what makes these two so perfect together. Leo provides the steady ground, while Priya brings the spark; they balance each other beautifully, making them an incredible team. I wish you both a lifetime of laughter, love, and many more adventures—though perhaps with slightly more seaworthy vessels.
Please join me in raising your glasses: To Priya and Leo!
Qwen · Qwen3.8 27B 💡 🏠
Hi everyone—Sam here. Priya and Leo have been my best friends since we were kids, so I’m honored to be here. One summer, the three of us built a raft out of whatever we could find. It was beautiful, creative, and completely leaky. We spent the afternoon bailing water while laughing like it was the greatest adventure ever. That’s kind of what their relationship feels like: joyful, a little messy sometimes, but always full of love. Priya brings warmth and wisdom; Leo brings steadiness and humor. Together they balance each other in the best way. So here’s to you two—may your future be filled with laughter, patience, adventure, and all the quiet moments that make a life together feel like home. Please raise your glasses.
