Gemma 4 31B vs Qwen3.8 27B
Compare Gemma 4 31B and Qwen3.8 27B side-by-side. See how these vision models stack up in Image Captioning, OCR, Open Prompt, Object Detection, and Classification.
Compare Gemma 4 31B vs Qwen3.8 27B live
Run the same image across every model that supports a task and compare their outputs side-by-side.
Detect and compare bounding boxes across models on the same image.
Upload an image
Drag and drop an image here, or click to browse
Models in this comparison
Gemma 4 31B vs Qwen3.8 27B on Vision Evals
Qwen3.8 27B scores higher on 5 of the six Vision Evals tasks.
The widest gap is Object Detection, where Qwen3.8 27B leads 65.7% to 47.5%.
Overall, Gemma 4 31B averages 67.0% (#36 of 61) against 74.7% (#22 of 61) for Qwen3.8 27B.
Qwen3.8 27B is both cheaper ($0.0009 vs $0.0015 per sample) and faster (18.0s vs 34.4s per sample).
Gemma 4 31B vs Qwen3.8 27B Comparison Table
Evals updated September 29, 2026Pricing updated September 29, 2026
| Property | Gemma 4 31B | Qwen3.8 27B |
|---|---|---|
| Organization | Qwen | |
| Category | open | open |
| Modality | multimodal | multimodal |
| Release Date | Apr 2026 | Aug 2026 |
| Context Window | 256K | 262K |
| Parameters | 31B | 27.78B |
| License | Apache 2.0 | Apache 2.0 |
| Pricing per 1M tokens | ||
| Input $/1M | $0.090 | $0.025 |
| Output $/1M | $0.340 | $4.35 |
| Vision Tasks | ||
| Captioning | Demo | Demo |
| Chart Question Answering | ||
| Classification | Demo | Demo |
| Document Question Answering | ||
| Image Tagging | ||
| Multi-Label Classification | ||
| Object Detection | Demo | Demo |
| OCR | Demo | Demo |
| Vision Language | ||
| Visual Question Answering | Demo | Demo |
| Model Features | ||
| Foundation Vision | ||
| LLMs with Vision Capabilities | ||
| Multimodal Vision | ||
Vision Evalsground-truth scores across 6 vision tasks, pooled at low effort | ||
| Overall | 67.0% | 74.7% |
| Quantizationsself-hosted | ||
| Avg cost / sample | $0.0015 | $0.0009 |
| Avg speed / sample | 34.36s | 17.99s |
| By task | ||
| Object Detection (low) | 47.5% ±0.6, Mean of 3 runs, range 46.8 to 48.0 | 65.7% ±1.0, Mean of 3 runs, range 64.6 to 66.5 |
| Object Detection (high) | – | 66.1% ±1.4, Mean of 3 runs, range 64.9 to 67.8 |
| Counting (low) | 51.4% ±2.7, Mean of 3 runs, range 48.6 to 54.0 | 64.9% ±4.1, Mean of 3 runs, range 60.8 to 68.9 |
| Counting (high) | – | 68.0% ±2.0, Mean of 3 runs, range 66.2 to 70.3 |
| Identification (low) | 79.2% ±1.6, Mean of 3 runs, range 78.1 to 81.3 | 85.4% ±4.7, Mean of 3 runs, range 81.3 to 90.6 |
| Identification (high) | – | 87.5% ±3.1, Mean of 3 runs, range 84.4 to 90.6 |
| OCR (low) | 90.8% ±0.6, Mean of 3 runs, range 90.2 to 91.5 | 92.2% ±1.2, Mean of 3 runs, range 91.1 to 93.4 |
| OCR (high) | – | 91.5% ±1.4, Mean of 3 runs, range 90.1 to 92.9 |
| Data Extraction (low) | 80.4% ±1.0, Mean of 3 runs, range 79.4 to 81.4 | 78.0% ±1.0, Mean of 3 runs, range 77.3 to 79.4 |
| Data Extraction (high) | – | 80.8% ±1.0, Mean of 3 runs, range 79.4 to 81.4 |
| Reasoning (low) | 52.8% ±1.3, Mean of 3 runs, range 51.7 to 54.3 | 62.0% ±2.0, Mean of 3 runs, range 60.3 to 64.2 |
| Reasoning (high) | – | 66.0% ±0.7, Mean of 3 runs, range 65.6 to 66.9 |
Gemma 4 31B vs Qwen3.8 27B: Overview
Gemma 4 31B is the largest dense model in Google's Gemma 4 family, built from the same research as Gemini 3 and released as open weights under the Apache 2.0 license. It supports a 256K token context window with text and image input, configurable thinking mode for step-by-step reasoning, and multilingual support across 140+ languages. The unquantized model fits on a single 80GB GPU.
For vision tasks, Gemma 4 31B supports image understanding with variable aspect ratios and resolutions, and can output structured bounding boxes for UI element detection, making it useful for document parsing and UI understanding. Compared to Gemma 3, it delivers stronger reasoning and multimodal performance. It is part of a four-size family alongside the 26B A4B MoE variant and two on-device models (E2B, E4B), with the 31B dense variant optimized for output quality and fine-tuning over inference speed.
Qwen3.8-27B is a dense vision-language model of roughly 27.78 billion parameters from Alibaba's Qwen team, built on the architectural foundation established by the Qwen3.5 generation. It takes interleaved text, image, and video input through an integrated vision encoder and produces text, covering document and chart understanding, visual question answering, front-end code generation from visual references, and agentic computer-use workflows. The decoder is a hybrid stack of 64 layers that repeats a pattern of three Gated DeltaNet linear-attention blocks, each followed by a feed-forward block, then one grouped-query full-attention block, so only a quarter of the layers maintain a conventional key-value cache. Native context length is 262,144 tokens.
Post-training targets long-horizon agentic work, and the model exposes per-request thinking control that lets callers switch reasoning traces on or off and choose a reasoning effort level. Multi-token prediction weights are trained alongside the main network for speculative decoding. Qwen reports improvements over Qwen3.6-27B on agentic and multimodal evaluations including Terminal-Bench 2.1, OSWorld-Verified, and SWE-bench Multimodal, with additional results on MathVision, CharXiv, and an internal Vision2Web suite.
Frequently Asked Questions
On Roboflow's Vision Evals, Qwen3.8 27B performed better. It scores higher on 5 of the six vision tasks and averages 74.7% (#22 of 61) against 67.0% (#36 of 61) for Gemma 4 31B. These scores measure vision capabilities only; either model may perform differently on work outside this benchmark's tasks.
No. On the Vision Evals Object Detection benchmark at low effort, Qwen3.8 27B leads with 65.7% against 47.5%. This is the widest gap between the two models across the benchmark's tasks.
Qwen3.8 27B is cheaper to run. Across Roboflow's Vision Evals task mix it averaged $0.0009 per sample against $0.0015. Actual costs depend on your image sizes, prompts, and output length.
Qwen3.8 27B is faster. Across Roboflow's Vision Evals it averaged 18.0s per inference against 34.4s. Latency varies with image size, prompt length, and provider load, so treat these as relative rather than guaranteed figures.
Yes. The comparison demo on this page runs both models on the same image side by side for image captioning and OCR in the free Roboflow Playground. You can try it instantly, and a free account unlocks unlimited runs.