Image understanding

Vision-language models ranked by how people rate their answers about images, the assistants you can show a picture or a screen, and local apps that run open vision models.

Models are ranked by LMArena Vision Arena (overall, style control), then Artificial Analysis MMMU-Pro (independent evaluation) for models without a result on the first. Tools inherit the tier of the best model they offer. How we rank.

Best overallSQwen3.8-Max1,301 on LMArena Vision Arena (overall, style control)
Best valueAMiMo-V2.6-FlashTier A for $0.18/M tokens
Best free startSOpenRouterFree tier · paid plans · via Qwen3.8-Max
Best local (≤24 GB GPU)AGemma 4 (31B)~19 GB VRAM · 1,262 on LMArena Vision Arena (overall, style control)

Tier: how close to the best result on the task's main public benchmark
Outlined: a tool inherits the tier of the best model it offers
No public result yet
Cloud · Local: runs on your own hardware

Models

The engines themselves. Use them through an API or inside the apps on the right.

Cloud

S Qwen3.8-MaxAlibaba (Qwen) · $3/M tokens1,301 on LMArena Vision Arena (overall, style control)$S

GPT-6.1 SolOpenAI · $4/M tokens1,294 on LMArena Vision Arena (overall, style control)S

Claude Opus 5.5Anthropic · $8/M tokens1,293 on LMArena Vision Arena (overall, style control)S Muse Spark 1.3Meta · $2/M tokens1,290 on LMArena Vision Arena (overall, style control)$S

Gemini 3.8 FlashGoogle · $1.5/M tokens1,289 on LMArena Vision Arena (overall, style control)$S

Claude Fable 5.1Anthropic · $20/M tokens1,286 on LMArena Vision Arena (overall, style control)S

GPT-6 AstraOpenAI · $20/M tokens1,286 on LMArena Vision Arena (overall, style control)S

Gemini 3.1 Pro (Preview)Google · $4.5/M tokens1,280 on LMArena Vision Arena (overall, style control)A

Claude Sonnet 5.5Anthropic · $4/M tokens1,266 on LMArena Vision Arena (overall, style control)A

Gemini 3.5 Flash-LiteGoogle · $0.85/M tokens1,266 on LMArena Vision Arena (overall, style control)$A

Qwen3.7-PlusAlibaba (Qwen) · $0.56/M tokens1,264 on LMArena Vision Arena (overall, style control)$A

Grok 4.7xAI · $3/M tokens1,248 on LMArena Vision Arena (overall, style control)A

GPT-6 LunaOpenAI · $0.2/M tokens1,241 on LMArena Vision Arena (overall, style control)$A

GLM-5V-TurboZ.ai · $1.9/M tokens1,231 on LMArena Vision Arena (overall, style control)B

Mistral Large 4 (preview)Mistral AI · $2.06/M tokens76.4% on Artificial Analysis MMMU-Pro (independent evaluation)

Local

GLM-5.3-FlashZ.ai · ~185 GB VRAM1,273 on LMArena Vision Arena (overall, style control)$A

Gemma 4 (31B)Google DeepMind · ~19 GB VRAM1,262 on LMArena Vision Arena (overall, style control)$A

DeepSeek-V4.1-Flash-1017DeepSeek · ~320 GB VRAM1,260 on LMArena Vision Arena (overall, style control)A

Step 5 PreviewStepFun · ~346 GB VRAM1,250 on LMArena Vision Arena (overall, style control)A

MiMo-V2.6-FlashXiaomi · ~178 GB VRAM1,248 on LMArena Vision Arena (overall, style control)A

MiMo-V2.6-ProXiaomi · ~585 GB VRAM1,246 on LMArena Vision Arena (overall, style control)A

Qwen3.8-27BAlibaba (Qwen) · ~17 GB VRAM1,244 on LMArena Vision Arena (overall, style control)$A

Muse Spark Open (816B-A42B)Meta · ~470 GB VRAM80.9% on Artificial Analysis MMMU-Pro (independent evaluation)A

Kimi K3Moonshot AI · ~1600 GB VRAM80.5% on Artificial Analysis MMMU-Pro (independent evaluation)A

MiniMax-M3MiniMax · ~245 GB VRAM1,236 on LMArena Vision Arena (overall, style control)A

Ling-3.0-flash-VLinclusionAI · ~73 GB VRAM79% on Artificial Analysis MMMU-Pro (independent evaluation)B

Inkling SmallThinking Machines Lab · ~159 GB VRAM1,208 on LMArena Vision Arena (overall, style control)B

Mistral Medium 3.5Mistral AI · ~75 GB VRAM1,200 on LMArena Vision Arena (overall, style control)B

Muse Glimmer 30BMeta · ~19 GB VRAM74.3% on Artificial Analysis MMMU-Pro (independent evaluation)B

InklingThinking Machines Lab · ~557 GB VRAM73.5% on Artificial Analysis MMMU-Pro (independent evaluation)B

Llama 4 MaverickMeta · ~230 GB VRAM1,145 on LMArena Vision Arena (overall, style control)C

Molmo2-8BAllen Institute for AI (Ai2) · ~6.5 GB VRAM1,112 on LMArena Vision Arena (overall, style control)$–

MiniCPM-V 4.7 (35B-A3B)OpenBMB · ~22 GB VRAMNo public result yet –

Moondream 3.1 (9B-A2B)Moondream (M87 Labs) · ~7 GB VRAMNo public result yet –

Nemotron 3.5 Bolt 74B-A9BNVIDIA · ~24 GB VRAMNo public result yet –

Nemotron 3.5 Super VL 120B-A12BNVIDIA · ~72 GB VRAMNo public result yet –

Qwen3.8-Flash-NextAlibaba (Qwen) · ~104 GB VRAMNo public result yet

Apps

Ready-made apps with a UI: sign in or install, then start.

Cloud

ChatGPTOpenAI · Free tier · $20/movia GPT-6.1 Sol S Microsoft CopilotMicrosoft · Free tier · $9.99/movia GPT-6.1 Sol$S

PerplexityPerplexity AI · Free tier · $20/movia GPT-6.1 Sol S

ClaudeAnthropic · Free tier · $20/movia Claude Opus 5.5 S

Meta AIMeta · Free tiervia Muse Spark 1.3 S

GeminiGoogle · Free tier · $4.99/movia Gemini 3.8 Flash$A

GrokSpaceXSI (formerly xAI) · Free tier · $8/movia Grok 4.7 A

KimiMoonshot AI · Free tiervia Kimi K3 B

Vibe (formerly Le Chat)Mistral AI · Free tier · $15/movia Mistral Medium 3.5 –

Be My EyesBe My Eyes · FreeNo public result yet –

EnvisionEnvision · FreeNo public result yet –

Google LensGoogleNo public result yet –

Qwen ChatAlibaba (Qwen team) · FreeNo public result yet – Seeing AIMicrosoft · FreeNo public result yet

Local

Open WebUIOpen WebUI Inc. · Freevia Gemma 4 (31B)$A

TextGen (text-generation-webui)oobabooga · Freevia Gemma 4 (31B)$A

JanMenlo Research · Freevia Gemma 4 (31B)$A

LM StudioElement Labs · Freevia Gemma 4 (31B)$A

Msty StudioMsty · Freevia Gemma 4 (31B)$

Runtimes & gateways

Run open models on your own machine, or reach many models through one API.

Cloud

OpenRouterOpenRouter · Free tiervia Qwen3.8-Max S

PerplexityPerplexity AI · Free tier · $20/movia GPT-6.1 Sol$

Local

OllamaOllama · Freevia GLM-5.3-Flash$A

llama.cppggml-org · Freevia Gemma 4 (31B)A

vLLMvLLM project · Freevia Gemma 4 (31B)A

LocalAIEttore Di Giacinto (mudler) · Freevia Gemma 4 (31B)A

SGLangSGLang project (LMSYS) · Freevia Gemma 4 (31B)A

LMDeployInternLM (Shanghai AI Laboratory) · Freevia Gemma 4 (31B)A

MLX-VLMBlaizzy (open-source project) · Freevia Gemma 4 (31B)

Quality vs. price

LMArena Vision Arena (overall, style control) score against API price (USD per 1M tokens, blended 3:1 input:output, log scale). Up and to the left is better; the line joins the models that nothing else beats on both.

Closed models
Open-weights models
Best-value frontier

Show as table

Model	Weights	LMArena Vision Arena (overall, style control)	Price	Frontier
Qwen3.8-Max	Closed	1,301	$3/M tokens	Yes
GPT-6.1 Sol	Closed	1,294	$4/M tokens
Claude Opus 5.5	Closed	1,293	$8/M tokens
Muse Spark 1.3	Closed	1,290	$2/M tokens	Yes
Gemini 3.8 Flash	Closed	1,289	$1.5/M tokens	Yes
Claude Fable 5.1	Closed	1,286	$20/M tokens
GPT-6 Astra	Closed	1,286	$20/M tokens
Gemini 3.1 Pro (Preview)	Closed	1,280	$4.5/M tokens
GLM-5.3-Flash	Open	1,273	$0.24/M tokens	Yes
Claude Sonnet 5.5	Closed	1,266	$4/M tokens
Gemini 3.5 Flash-Lite	Closed	1,266	$0.85/M tokens
Qwen3.7-Plus	Closed	1,264	$0.56/M tokens
Gemma 4 (31B)	Open	1,262	$0.21/M tokens	Yes
DeepSeek-V4.1-Flash-1017	Open	1,260	$0.52/M tokens
Step 5 Preview	Open	1,250	$1.43/M tokens
MiMo-V2.6-Flash	Open	1,248	$0.18/M tokens	Yes
Grok 4.7	Closed	1,248	$3/M tokens
MiMo-V2.6-Pro	Open	1,246	$0.54/M tokens
Qwen3.8-27B	Open	1,244	$0.96/M tokens
GPT-6 Luna	Closed	1,241	$0.2/M tokens
MiniMax-M3	Open	1,236	$0.52/M tokens
GLM-5V-Turbo	Closed	1,231	$1.9/M tokens
Inkling Small	Open	1,208	$0.64/M tokens
Mistral Medium 3.5	Open	1,200	$3/M tokens
Llama 4 Maverick	Open	1,145	$0.55/M tokens

Everything, in order

All image understanding options in quality order
Tier	Name	Approach	Runs	Rated by	Price	Skill
S	Qwen3.8-Max Alibaba (Qwen)	Models	Cloud	1,301 on LMArena Vision Arena (overall, style control)	$3/M tokens	—
S	GPT-6.1 Sol OpenAI	Models	Cloud	1,294 on LMArena Vision Arena (overall, style control)	$4/M tokens	—
S	Claude Opus 5.5 Anthropic	Models	Cloud	1,293 on LMArena Vision Arena (overall, style control)	$8/M tokens	—
S	Muse Spark 1.3 Meta	Models	Cloud	1,290 on LMArena Vision Arena (overall, style control)	$2/M tokens	—
S	Gemini 3.8 Flash Google	Models	Cloud	1,289 on LMArena Vision Arena (overall, style control)	$1.5/M tokens	—
S	Claude Fable 5.1 Anthropic	Models	Cloud	1,286 on LMArena Vision Arena (overall, style control)	$20/M tokens	—
S	GPT-6 Astra OpenAI	Models	Cloud	1,286 on LMArena Vision Arena (overall, style control)	$20/M tokens	—
S	Gemini 3.1 Pro (Preview) Google	Models	Cloud	1,280 on LMArena Vision Arena (overall, style control)	$4.5/M tokens	—
S	GLM-5.3-Flash Z.ai	Models	Local	1,273 on LMArena Vision Arena (overall, style control)	~185 GB VRAM	—
S	OpenRouter OpenRouter	Runtimes & gateways	Cloud	via Qwen3.8-Max	Free tier	intermediate
S	ChatGPT OpenAI	Apps	Cloud	via GPT-6.1 Sol	Free tier · $20/mo	beginner
S	Microsoft Copilot Microsoft	Apps	Cloud	via GPT-6.1 Sol	Free tier · $9.99/mo	beginner
S	Perplexity Perplexity AI	Apps + Runtimes & gateways	Cloud	via GPT-6.1 Sol	Free tier · $20/mo	beginner
S	Claude Anthropic	Apps	Cloud	via Claude Opus 5.5	Free tier · $20/mo	beginner
S	Meta AI Meta	Apps	Cloud	via Muse Spark 1.3	Free tier	beginner
S	Gemini Google	Apps	Cloud	via Gemini 3.8 Flash	Free tier · $4.99/mo	beginner
S	Ollama Ollama	Runtimes & gateways	Local	via GLM-5.3-Flash	Free	intermediate
A	Claude Sonnet 5.5 Anthropic	Models	Cloud	1,266 on LMArena Vision Arena (overall, style control)	$4/M tokens	—
A	Gemini 3.5 Flash-Lite Google	Models	Cloud	1,266 on LMArena Vision Arena (overall, style control)	$0.85/M tokens	—
A	Qwen3.7-Plus Alibaba (Qwen)	Models	Cloud	1,264 on LMArena Vision Arena (overall, style control)	$0.56/M tokens	—
A	Gemma 4 (31B) Google DeepMind	Models	Local	1,262 on LMArena Vision Arena (overall, style control)	~19 GB VRAM	—
A	DeepSeek-V4.1-Flash-1017 DeepSeek	Models	Local	1,260 on LMArena Vision Arena (overall, style control)	~320 GB VRAM	—
A	Step 5 Preview StepFun	Models	Local	1,250 on LMArena Vision Arena (overall, style control)	~346 GB VRAM	—
A	Grok 4.7 xAI	Models	Cloud	1,248 on LMArena Vision Arena (overall, style control)	$3/M tokens	—
A	MiMo-V2.6-Flash Xiaomi	Models	Local	1,248 on LMArena Vision Arena (overall, style control)	~178 GB VRAM	—
A	MiMo-V2.6-Pro Xiaomi	Models	Local	1,246 on LMArena Vision Arena (overall, style control)	~585 GB VRAM	—
A	Qwen3.8-27B Alibaba (Qwen)	Models	Local	1,244 on LMArena Vision Arena (overall, style control)	~17 GB VRAM	—
A	GPT-6 Luna OpenAI	Models	Cloud	1,241 on LMArena Vision Arena (overall, style control)	$0.2/M tokens	—
A	Muse Spark Open (816B-A42B) Meta	Models	Local	80.9% on Artificial Analysis MMMU-Pro (independent evaluation)	~470 GB VRAM	—
A	Kimi K3 Moonshot AI	Models	Local	80.5% on Artificial Analysis MMMU-Pro (independent evaluation)	~1600 GB VRAM	—
A	MiniMax-M3 MiniMax	Models	Local	1,236 on LMArena Vision Arena (overall, style control)	~245 GB VRAM	—
A	GLM-5V-Turbo Z.ai	Models	Cloud	1,231 on LMArena Vision Arena (overall, style control)	$1.9/M tokens	—
A	Ling-3.0-flash-VL inclusionAI	Models	Local	79% on Artificial Analysis MMMU-Pro (independent evaluation)	~73 GB VRAM	—
A	Open WebUI Open WebUI Inc.	Apps	Local	via Gemma 4 (31B)	Free	intermediate
A	llama.cpp ggml-org	Runtimes & gateways	Local	via Gemma 4 (31B)	Free	advanced
A	vLLM vLLM project	Runtimes & gateways	Local	via Gemma 4 (31B)	Free	advanced
A	LocalAI Ettore Di Giacinto (mudler)	Runtimes & gateways	Local	via Gemma 4 (31B)	Free	advanced
A	TextGen (text-generation-webui) oobabooga	Apps	Local	via Gemma 4 (31B)	Free	intermediate
A	Jan Menlo Research	Apps	Local	via Gemma 4 (31B)	Free	beginner
A	SGLang SGLang project (LMSYS)	Runtimes & gateways	Local	via Gemma 4 (31B)	Free	advanced
A	LMDeploy InternLM (Shanghai AI Laboratory)	Runtimes & gateways	Local	via Gemma 4 (31B)	Free	advanced
A	MLX-VLM Blaizzy (open-source project)	Runtimes & gateways	Local	via Gemma 4 (31B)	Free	advanced
A	LM Studio Element Labs	Apps	Local	via Gemma 4 (31B)	Free	beginner
A	Msty Studio Msty	Apps	Local	via Gemma 4 (31B)	Free	beginner
A	Grok SpaceXSI (formerly xAI)	Apps	Cloud	via Grok 4.7	Free tier · $8/mo	beginner
A	Kimi Moonshot AI	Apps	Cloud	via Kimi K3	Free tier	beginner
B	Inkling Small Thinking Machines Lab	Models	Local	1,208 on LMArena Vision Arena (overall, style control)	~159 GB VRAM	—
B	Mistral Medium 3.5 Mistral AI	Models	Local	1,200 on LMArena Vision Arena (overall, style control)	~75 GB VRAM	—
B	Mistral Large 4 (preview) Mistral AI	Models	Cloud	76.4% on Artificial Analysis MMMU-Pro (independent evaluation)	$2.06/M tokens	—
B	Muse Glimmer 30B Meta	Models	Local	74.3% on Artificial Analysis MMMU-Pro (independent evaluation)	~19 GB VRAM	—
B	Inkling Thinking Machines Lab	Models	Local	73.5% on Artificial Analysis MMMU-Pro (independent evaluation)	~557 GB VRAM	—
B	Llama 4 Maverick Meta	Models	Local	1,145 on LMArena Vision Arena (overall, style control)	~230 GB VRAM	—
B	Vibe (formerly Le Chat) Mistral AI	Apps	Cloud	via Mistral Medium 3.5	Free tier · $15/mo	beginner
C	Molmo2-8B Allen Institute for AI (Ai2)	Models	Local	1,112 on LMArena Vision Arena (overall, style control)	~6.5 GB VRAM	—
–	Be My Eyes Be My Eyes	Apps	Cloud	No public result yet	Free	beginner
–	Envision Envision	Apps	Cloud	No public result yet	Free	beginner
–	Google Lens Google	Apps	Cloud	No public result yet	—	beginner
–	MiniCPM-V 4.7 (35B-A3B) OpenBMB	Models	Local	No public result yet	~22 GB VRAM	—
–	Moondream 3.1 (9B-A2B) Moondream (M87 Labs)	Models	Local	No public result yet	~7 GB VRAM	—
–	Nemotron 3.5 Bolt 74B-A9B NVIDIA	Models	Local	No public result yet	~24 GB VRAM	—
–	Nemotron 3.5 Super VL 120B-A12B NVIDIA	Models	Local	No public result yet	~72 GB VRAM	—
–	Qwen Chat Alibaba (Qwen team)	Apps	Cloud	No public result yet	Free	beginner
–	Qwen3.8-Flash-Next Alibaba (Qwen)	Models	Local	No public result yet	~104 GB VRAM	—
–	Seeing AI Microsoft	Apps	Cloud	No public result yet	Free	beginner

Benchmarks used

LMArena Vision Arena (overall, style control) by LMArena (arena.ai) (blind human votes, Arena score (Bradley-Terry rating from pairwise human votes on image + text prompts, style-controlled view)) · snapshot 9 Oct 2026
Read from LMArena's official Hugging Face dataset (vision_style_control, latest split, category overall). 165 entries; covers nearly all current closed frontier models and the main open families (Gemma 4, Qwen3.8-27B, GLM-5.3-Flash, DeepSeek-V4.1-Flash, MiMo-V2.6-Pro, MiniMax-M3, Mistral Medium 3.5). Not listed: Gemini 4 Argon, Claude Haiku 5.5, Kimi K3, Muse Wonder, GLM-5.3, DeepSeek-V4-Pro. The #1 entry is the older Claude Fable 5; several current entries have under 3k votes.
Artificial Analysis MMMU-Pro (independent evaluation) by Artificial Analysis (automated tests, Accuracy on MMMU-Pro (multi-discipline multimodal questions over images), as run by Artificial Analysis) · snapshot 19 Oct 2026
Values read from the model data embedded in https://artificialanalysis.ai/leaderboards/models (124 model variants with an MMMU-Pro result; the evaluation page itself only embeds its 19 chart defaults). No last-updated date shown. Covers most current closed frontier models and many open ones (Gemma 4, Qwen3.8-27B, Kimi K3, MiniMax-M3, DeepSeek V4.1 Flash, Muse Spark Open/Glimmer, Inkling, Step 5, Mistral Large 4, MiMo-V2.6-Flash). Not listed: Claude Fable 5.1, Claude Sonnet 5.5, Claude Haiku 5.5, Gemini 4 Argon, GLM-5.3/5.3-Flash, MiMo-V2.6-Pro, MiniCPM-V 4.7 (only 4.6 1.3B). Several entries are effort variants of one model; record the best. Independently run (not vendor-reported), unlike aggregator MMMU tables.

Image understanding

Quality vs. price

Quality vs. setup

Everything, in order

Benchmarks used