Roboflow

Grok 4.7 vs Qwen3.6 35B A3B

Compare Grok 4.7 and Qwen3.6 35B A3B side-by-side.

Compare Grok 4.7 vs Qwen3.6 35B A3B live

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

These models don't share enough common tasks for a side-by-side demo. See the comparison table below for their capabilities.

Models in this comparison

Grok 4.7 vs Qwen3.6 35B A3B on Vision Evals

Grok 4.7 scores higher on 4 of the six Vision Evals tasks.

The widest gap is Object Detection, where Qwen3.6 35B A3B leads 57.0% to 40.4%.

Overall, Grok 4.7 averages 71.9% (#22 of 54) against 71.9% (#21 of 54) for Qwen3.6 35B A3B.

Qwen3.6 35B A3B is cheaper ($0.0012 vs $0.012 per sample), while Grok 4.7 is faster (23.6s vs 27.1s per sample).

Grok 4.7Qwen3.6 35B A3B

Grok 4.7 vs Qwen3.6 35B A3B Comparison Table

Evals updated September 22, 2026Pricing updated September 22, 2026

PropertyGrok 4.7Qwen3.6 35B A3B
OrganizationSpaceXAIQwen
Categoryclosedopen
Modalitymultimodal
Release DateSep 2026Apr 2026
Context Window500K262K
Parameters35B total, 3B active
LicenseProprietaryApache 2.0
Pricing per 1M tokens
Input $/1M$1.60$0.150
Output $/1M$4.80$1.00
Vision Tasks
CaptioningDemo
Chart Question Answering
ClassificationDemo
Document Question Answering
Image Tagging
Multi-Label Classification
Object DetectionDemo
OCRDemo
Vision Language
Visual Question AnsweringDemo
Phrase Grounding
Video Classification
Model Features
Foundation Vision
LLMs with Vision Capabilities
Multimodal Vision
Vision Evalsground-truth scores across 6 vision tasks, pooled at low effort
Overall
71.9%
71.9%
Quantizationsself-hosted
FP871.9%AWQ-INT471.3%hardware →
Avg cost / sample$0.012$0.0012
Avg speed / sample23.55s27.10s
By task
Object Detection (low)
40.4%
±0.6, Mean of 3 runs, range 39.8 to 41.0
$0.017
57.0%
±1.3, Mean of 3 runs, range 56.1 to 58.7
$0
Object Detection (high)
41.2%
±1.6, Mean of 3 runs, range 39.6 to 42.8
$0.023
Counting (low)
61.7%
±1.3, Mean of 3 runs, range 60.8 to 63.5
$0.0086
65.3%
±2.7, Mean of 3 runs, range 62.2 to 67.6
$0
Counting (high)
60.8%
±1.3, Mean of 3 runs, range 59.5 to 62.2
$0.013
Identification (low)
87.5%
±3.1, Mean of 3 runs, range 84.4 to 90.6
$0.0050
82.3%
±6.3, Mean of 3 runs, range 75.0 to 87.5
$0
Identification (high)
80.2%
±1.6, Mean of 3 runs, range 78.1 to 81.3
$0.0074
OCR (low)
92.6%
±0.7, Mean of 3 runs, range 92.1 to 93.4
$0.014
87.7%
±0.0, Mean of 3 runs, range 87.6 to 87.7
$0
OCR (high)
93.5%
±0.3, Mean of 3 runs, range 93.1 to 93.8
$0.034
Data Extraction (low)
84.9%
±2.6, Mean of 3 runs, range 82.5 to 87.6
$0.0048
84.5%
±1.0, Mean of 3 runs, range 83.5 to 85.6
$0
Data Extraction (high)
87.6%
±1.5, Mean of 3 runs, range 86.6 to 89.7
$0.0054
Reasoning (low)
64.2%
±2.3, Mean of 3 runs, range 62.3 to 66.9
$0.012
54.8%
±1.3, Mean of 3 runs, range 53.0 to 55.6
$0
Reasoning (high)
66.9%
±1.3, Mean of 3 runs, range 65.6 to 68.2
$0.019

Grok 4.7 vs Qwen3.6 35B A3B: Overview

Grok 4.7

Grok 4.7 is a proprietary model from SpaceXAI, released on September 21, 2026. It accepts text and images as input and returns text. It extends Grok 4.6 and is listed at the same API price.

Its Vision Evals scores are on the leaderboard. Running it in the Playground is not available yet, because the inference workflow is not ready.

Qwen3.6 35B A3B

Qwen3.6-35B-A3B is a sparse Mixture-of-Experts (MoE) multimodal language model developed by the Qwen team at Alibaba Group. It carries 35 billion total parameters but activates only approximately 3 billion per forward pass via a learned routing mechanism, giving it the representational capacity of a large dense model at a fraction of the inference compute. The model is natively multimodal, processing images, documents, and video alongside text as a core architectural capability rather than an add-on. It supports a native context window of 262,144 tokens, extensible up to 1,010,000 tokens via YaRN. A key design feature is the unified thinking/non-thinking mode framework: users can switch between deliberate chain-of-thought reasoning and fast direct responses within a single model, and a "thinking preservation" option retains reasoning context across multi-turn agentic workflows to reduce redundant computation.

The model is specifically optimized for agentic coding tasks, including repository-level reasoning, frontend workflow generation, multi-step tool use, and MCP (Model Context Protocol) integration. On SWE-bench Verified it scores 73.4%, on Terminal-Bench 2.0 it scores 51.5%, and on MCPMark it scores 37.0%. For vision-language tasks it achieves 92.0 on RefCOCO, 89.9 on OmniDocBench 1.5, and 83.7 on VideoMMMU. The model also supports Multi-Token Prediction (MTP) for speculative decoding. All Qwen3.6 open-weight models are released under the Apache 2.0 license.