Gemma 4 31B vs Qwen3.8 27B for local text
Two dense ~30B models for a 24 to 48 GB machine
Facts from the reviewed catalog
| Specification | Gemma 4 31B | Qwen3.8 27B |
|---|---|---|
| Task | Text & chat | Text & chat |
| Parameters | 30.7B | 27B |
| License | Apache 2.0 | Apache 2.0 |
| Commercial use | Allowedlicense text | Allowedlicense text |
| Quality · LMArena 2026-10-22 · text | 1,452 ± 71,445–1,460 · 6,137 votes · rank 76 · source | 1,437 ± 51,432–1,442 · 26,238 votes · rank 102 · source |
| Memory | Q4 planning estimate (~24 GB of weights before the machine's reserve) | Q4 planning estimate (~21 GB of weights before the machine's reserve) |
| Checkpoint | official card | official card |
In LMArena 2026-10-22 · text, Gemma 4 31B rates 1,452 ± 7 against 1,437 ± 5 for Qwen3.8 27B, 15 points ahead with intervals that do not overlap. Blind pairwise votes by people on the hosted model, CC BY 4.0. A rating is relative; overlapping intervals mean the order is not settled. Scale: Arena rating (style control, overall) ± half the confidence interval.
Memory on three machines
Each machine's own reserve is added (graphics cards 2 GB; unified and system memory 25%, 2 to 8 GB). Planning estimates, not measured runs.
| Machine | Gemma 4 31B | Qwen3.8 27B |
|---|---|---|
| RTX 4090 · 24 GB VRAM | ~26 GBIn memory, outside the fast part · offload to PC RAM (32 GB assumed) | ~23 GBFast memory holds it |
| RTX 5090 · 32 GB VRAM | ~26 GBFast memory holds it | ~23 GBFast memory holds it |
| Mac mini M4 Pro · 48 GB unified | ~32 GBFast memory holds it | ~29 GBFast memory holds it |
Gemma 4 31B: Q4 planning estimate (~24 GB of weights before the machine's reserve). Qwen3.8 27B: Q4 planning estimate (~21 GB of weights before the machine's reserve). Formulas
Everything that runs on each machine: RTX 4090 · 24 GB VRAM · RTX 5090 · 32 GB VRAM · Mac mini M4 Pro · 48 GB unified
Which to choose
Two dense models of similar size: every token reads all of the weights, so speed on a given machine is similar and set by memory bandwidth. Gemma 4 31B (30.7 billion) needs about 24 GB of Q4 weights, Qwen3.8 27B about 21 GB, so on a 24 GB card only Qwen3.8 stays in the fast memory. Both are Apache 2.0 with a 262,144-token context and read images; Qwen3.8 also takes video. Their LMArena text ratings are 14 points apart, which is close on this scale; read the intervals below before calling either one better.
Files and run commands
The default file of the most downloaded collected source, or the commands each card documents. Not run by this site.
Gemma 4 31B
unsloth/gemma-4-31B-it-GGUF · gemma-4-31B-it-Q4_K_M.gguf · Q4_K_M · 18.3 GB · unsloth
ollama run hf.co/unsloth/gemma-4-31B-it-GGUF:Q4_K_Mllama-server -hf unsloth/gemma-4-31B-it-GGUF:Q4_K_Mlms get https://huggingface.co/unsloth/gemma-4-31B-it-GGUF- mlx-lm: mlx-lm runs MLX repositories; pick the MLX version on a Mac
Qwen3.8 27B
unsloth/Qwen3.8-27B-GGUF · Qwen3.8-27B-UD-Q4_K_M.gguf · UD-Q4_K_M · 16.5 GB · unsloth
ollama run hf.co/unsloth/Qwen3.8-27B-GGUF:UD-Q4_K_Mllama-server -hf unsloth/Qwen3.8-27B-GGUF:UD-Q4_K_Mlms get https://huggingface.co/unsloth/Qwen3.8-27B-GGUF- mlx-lm: mlx-lm runs MLX repositories; pick the MLX version on a Mac
Numbers each publisher reports
Each card's own results with its own settings. Publishers choose different tests and settings, so these values are never compared across publishers here.
Gemma 4 31B · reported by Google
- MMLU Pro
- 85.2%
- AIME 2026 no tools
- 89.2%
- LiveCodeBench v6
- 80.0%
- Codeforces ELO
- 2,150
- GPQA Diamond
- 84.3%
- Tau2 (average over 3)
- 76.9%
- HLE no tools
- 19.5%
- HLE with search
- 26.5%
Every reported result
Values as printed on the card, with the publisher's own settings (card section: Benchmark Results); column “Gemma 4 31B”. Not comparable across publishers and not used in any ranking. Model card
Qwen3.8 27B · reported by Alibaba
- Agentic terminal coding Terminal Bench 2.1 (Terminus)
- 73.0Text Performance
- Agentic coding SWE-bench Pro
- 61.7Text Performance
- Repo-level code generation NL2Repo-Bench
- 42.3Text Performance
- Agentic coding DeepSWE 1.1
- 42.2Text Performance
- Software engineering QwenSWEBench
- 79.0Text Performance
- Long-horizon office work CoWorkBench
- 70.7Text Performance
- Professional job tasks JobBench
- 33.4Text Performance
- Instruction following IFBench
- 79.5Text Performance
Every reported result
- Scientific reasoning GPQA Diamond
- 89.2Text Performance
- Multidisciplinary reasoning HLE
- 30.8Text Performance
- Competitive coding LiveCodeBench v6
- 90.3Text Performance
- Computer use OSWorld-Verified
- 84.3VL Performance
- Browser use WebArena-Verified
- 64.8VL Performance
- Mobile use AndroidWorld
- 81.9VL Performance
- Application recreation RecreationBench
- 47.1VL Performance
- Multimodal software engineering SWE-MM
- 38.6VL Performance
- Visual web development Vision2Web
- 62.9VL Performance
- Document intelligence OmniDocBench 1.5
- 91.1VL Performance
- Real-world perception RealWorldQA
- 85.9VL Performance
- Embodied intelligence ERQA
- 65.5VL Performance
Values as printed on the card, with the publisher's own settings (card section: Benchmark Results › Text Performance; Benchmark Results › VL Performance); column “Qwen3.8-27B”. Not comparable across publishers and not used in any ranking. Model card
Limits and dates
The published quality benchmark does not test a local quantized build. Memory estimates are not speed measurements or guaranteed hardware fits. Repository size, when available, is not VRAM. Formulas
Page updated · Catalog reviewed · LMArena 2026-10-22 · text