Roboflow

Gemma 3 27B vs Mistral Large 4

Mistral Large 4: 56.8% average across five Vision Evals tasks (#51 of 73 models). We have not run Gemma 3 27B through Roboflow's Vision Evals, so there is no head-to-head score. Run both on your own image below.

Compare Gemma 3 27B vs Mistral Large 4 live

Run the same image across every model that supports a task and compare their outputs side-by-side.

Extract and compare text from images across multiple models.

Open OCR in the full playground
GoogleGemma 3 27B
Run to compare this model.
MistralMistral Large 4
Run to compare this model.

Gemma 3 27B vs Mistral Large 4 Comparison Table

Evals updated November 19, 2026Pricing updated November 22, 2026

PropertyGemma 3 27BMistral Large 4
OrganizationGoogleMistral
Categoryopenopen
Modalitymultimodalmultimodal
Release DateMar 2025Oct 2026
Context Window128K1.0M
ParametersUnknown1.05T total, 49B active
LicenseCustomCustom
Commercial useSee license termsSee license terms
Pricing per 1M tokens
Input $/1M$0.080$0.700
Output $/1M$0.450$2.10
Vision Tasks
ClassificationSupportedDemo
Object DetectionNot listedDemo
Phrase GroundingNot listedSupported
Both support 8 more tasks
Model Features
Both support 3 more features
Vision Evalsground-truth scores across 5 vision tasks, pooled at low effort
OverallNot evaluated
56.8%
Avg cost / sample–$0.0033
Avg speed / sample–31.23s
By task
Object Detection (low)–
59.3%
±0.7, Mean of 3 runs, range 58.5 to 60.0
$0.0028
Object Detection (high)–
50.2%
±2.5, Mean of 3 runs, range 48.0 to 53.0
$0.023
Counting (low)–
54.5%
±0.7, Mean of 3 runs, range 54.0 to 55.4
$0.0011
Counting (high)–
63.1%
±2.0, Mean of 3 runs, range 60.8 to 64.9
$0.0095
Identification (low)–
85.4%
±3.1, Mean of 3 runs, range 81.3 to 87.5
$0.0009
Identification (high)–
85.4%
±1.6, Mean of 3 runs, range 84.4 to 87.5
$0.0044
OCR (low)–
46.1%
$0.0044
by category
Single value
33.5%
Transcription
68.9%
Structured JSON
71.7%
Text localization
0.0%
OCR (high)–
54.5%
$0.021
by category
Single value
38.3%
Transcription
75.3%
Structured JSON
76.4%
Text localization
28.8%
Reasoning (low)–
38.9%
±0.3, Mean of 3 runs, range 38.4 to 39.1
$0.0013
Reasoning (high)–
57.6%
±2.0, Mean of 3 runs, range 55.6 to 59.6
$0.013

Gemma 3 27B vs Mistral Large 4: Overview

Gemma 3 27B

Gemma 3 27B, announced on March 12, 2025, is the largest open-weight model in Google DeepMind’s Gemma 3 family. With around 27 billion parameters, it is multimodal—accepting both text and images as input and producing text outputs. It supports a 128,000-token context window and typically generates up to ~8,192 tokens, enabling it to process multi-page documents, extended conversations, or large batches of images in a single prompt.

The model is instruction-tuned in its “-it” variants for chat, reasoning, and summarization use cases, and it supports structured outputs and function calling. It is multilingual, covering over 140 languages. Deployment is flexible: the full BF16 model requires ~46 GB of VRAM, but quantization-aware training (QAT) versions in 8-bit or 4-bit reduce the footprint significantly, allowing more accessible use outside large-scale clusters. While it delivers stronger reasoning and multimodal performance than smaller Gemma models, it remains lighter and more open than proprietary systems, making it well-suited for research, development, and fine-tuned applications.

Mistral Large 4

Mistral Large 4, nicknamed Le Chonk, is a natively multimodal mixture-of-experts model from Mistral that accepts interleaved text and image input and produces text output. It uses a granular MoE design with roughly 1.05 trillion total parameters and 49 billion active per token, reported as 52 billion when embeddings and output layers are counted, paired with a 1.6 billion parameter vision encoder and a context window of one million tokens. The model is trained from scratch on about 3,800 NVIDIA Grace Blackwell GPUs in Mistral's European data centers and supports more than 160 languages. It behaves as a hybrid instruct and reasoning system, with a reasoning effort setting that selects between direct answers and longer deliberation, alongside function calling and structured output for agentic workflows.

Image understanding is a focus of this generation, covering documents, charts, technical drawings and natural scenes, and the model emits bounding box coordinates for visual grounding queries. Reported grounding results include 42 percent on Dense200 and 73 percent on the DIOR-RSVG remote sensing benchmark. Mistral describes agentic vision workflows in which the model zooms into gigapixel satellite imagery or engineering drawings to verify details, and reports coding results such as 62 percent on DeepSWE. Figures published at preview time are preliminary because the reinforcement learning phase is still in progress.