whatmodel.sh · find the best AI models you can run

Best AI models for images and documents to run locally on a 32GB Mac

A leaderboard of the best AI models for images and documents you can run locally on a Mac with 32GB of memory. Qwen3.8 27B is #1 right now, followed by Gemma 4 31B and Qwen3.6 27B. Models are ranked on tests that ask questions about images, charts and documents, like MMMU and CharXiv. Each model shows the download that fits and the command to run it.

38 models · scores combine 47 benchmarks · updated how it works · find a leaderboard · benchmark your machine
task[all tasks][agent tasks][coding][chat & writing][math][vision & documents][web search]

Scores use reviewed vision & documents evaluations only. Compare scores within this task. Models without task measurements are excluded.

Models for 32GB RAM · Apple Silicon → ~21GB fast · local only · reset
+ more
chip[M1 Pro][M1 Max][M2 Pro][M2 Max][M4][M5][M6 (24/32GB)]
set[all sets][Arena vision (style controlled) · 199][CharXiv descriptive questions · 2]

Filter the model list while keeping task scores. Models without a shared score scale are not yet calibrated; select a set to see its own scores.

Detect your exact chip or graphics card:

curl -G whatmodel.sh/ -d max=0 --data-urlencode "mac=$(sysctl -n hw.memsize machdep.cpu.brand_string)"Mac

curl -G whatmodel.sh/ -d max=0 --data-urlencode "ram=$(awk '/MemT/{print $2}' /proc/meminfo)" --data-urlencode "nv=$(nvidia-smi --query-gpu=name,memory.total --format=csv,noheader 2>/dev/null)"Linux

curl.exe -G whatmodel.sh/ -d max=0 --data-urlencode "ram=$((gcim Win32_ComputerSystem).TotalPhysicalMemory)" --data-urlencode "nv=$(nvidia-smi --query-gpu=name,memory.total --format=csv,noheader 2>$null)"Windows

$ terminal

detecting your OS…

Run curl "whatmodel.sh/?ram=32&gpu=apple&max=0"to get this list in your terminal (add &format=json for agents).

ram[8][16][24][32][48][64][96][128][192][256]GB
gpu[apple silicon][graphics card][none]
price[local][$0.05][$0.25][$1][$5][$20][any]max per M

detecting your OS…

Score vs memory to run

#modeltask score↓where↓speed↓RAM↓context↓get itsources
#modeltask score↓where↓speed↓RAM↓context↓get itsources
42Qwen3.8 27B · 4-bit Alibaba64±2local~19GB262kOllama › LMArena leaderboard datasetQwen (Alibaba) vendor-reported benchmark scores for its own models (qwen.ai launch posts, Qwen2.5-VL and Qwen3-VL technical reports, Qwen3.8-Flash-Next card)Qwen/Qwen3.8-27B model cardSciDocBench: A Workflow-Centered Benchmark and Data Pipeline for Scientific Document Understanding (arXiv:2609.05141v2)How Far Is Document Parsing from Solved? PureDocBench: A Source-Traceable Benchmark across Clean, Degraded, and Real-World Settings (arXiv:2605.07492v2)
52Gemma 4 31B · 4-bit Google62±2local~21GB262kOllama › LMArena leaderboard datasetGemma Team — Gemma 4 Technical Reportmeta-models/Muse-Glimmer-30B model card (speed and benchmark tables)ChartArena: Benchmarking Chart Parsing across Languages, Scenarios, and Formats (arXiv:2606.01348v3)BEAR-Bench: A Bilingual Enterprise and Academic Reasoning Benchmark for Multimodal Models (arXiv:2608.17895v1)
53Qwen3.6 27B · 4-bit Alibaba62±2local~20GB262kOllama › Qwen/Qwen3.6-27B model cardQwen/Qwen3.8-27B model cardmeta-models/Muse-Glimmer-30B model card (speed and benchmark tables)How Far Is Document Parsing from Solved? PureDocBench: A Source-Traceable Benchmark across Clean, Degraded, and Real-World Settings (arXiv:2605.07492v2)
61Gemma 4 26B-A4B · 4-bit Google61±2local~19GB262kllama.cpp › LMArena leaderboard datasetGemma Team — Gemma 4 Technical Reportgoogle/diffusiongemma-26B-A4B-it model card
64Qwen3.5 27B · 4-bit Alibaba60±1local~20GB262kOllama › Qwen/Qwen3.5-122B-A10B model cardLMArena leaderboard datasetQwen/Qwen3.6-35B-A3B model cardQwen/Qwen3.5-27B model cardSpaceCast-Bench: Evaluating Predictive Spatial Reasoning in Vision-Language Models (arXiv:2610.12402v1)BEAR-Bench: A Bilingual Enterprise and Academic Reasoning Benchmark for Multimodal Models (arXiv:2608.17895v1)+1
68Muse Glimmer 30B · 4-bit Meta60±4†local~21GB131kOllama › meta-models/Muse-Glimmer-30B model card (speed and benchmark tables)
78Qwen3.5 9B · 4-bit Alibaba57±2local~12GB262kLM Studio › Qwen/Qwen3.5-9B model cardChartArena: Benchmarking Chart Parsing across Languages, Scenarios, and Formats (arXiv:2606.01348v3)Qwen/Qwen3.5-4B model cardBEAR-Bench: A Bilingual Enterprise and Academic Reasoning Benchmark for Multimodal Models (arXiv:2608.17895v1)SpaceCast-Bench: Evaluating Predictive Spatial Reasoning in Vision-Language Models (arXiv:2610.12402v1)How Far Is Document Parsing from Solved? PureDocBench: A Source-Traceable Benchmark across Clean, Degraded, and Real-World Settings (arXiv:2605.07492v2)
80Gemma 4 12B · 8-bit Google56±4†local~15GB262kllama.cpp › Gemma Team — Gemma 4 Technical Report
97Qwen3.5 4B · 4-bit Alibaba53±2local~6.9GB262kLM Studio › Qwen/Qwen3.5-9B model cardChartArena: Benchmarking Chart Parsing across Languages, Scenarios, and Formats (arXiv:2606.01348v3)Qwen/Qwen3.5-4B model cardBEAR-Bench: A Bilingual Enterprise and Academic Reasoning Benchmark for Multimodal Models (arXiv:2608.17895v1)SpaceCast-Bench: Evaluating Predictive Spatial Reasoning in Vision-Language Models (arXiv:2610.12402v1)How Far Is Document Parsing from Solved? PureDocBench: A Source-Traceable Benchmark across Clean, Degraded, and Real-World Settings (arXiv:2605.07492v2)
107Apriel 1.6 15B Thinker · 8-bit ServiceNow50±4*†local~21GBllama.cpp › ServiceNow-AI/Apriel-1.6-15b-Thinker model card
110Qwen3-VL 30B-A3B Thinking · 4-bit Alibaba49±2†local~21GB256kllama.cpp › Qwen (Alibaba) vendor-reported benchmark scores for its own models (qwen.ai launch posts, Qwen2.5-VL and Qwen3-VL technical reports, Qwen3.8-Flash-Next card)
115Qwen3-VL 8B Thinking · 4-bit Alibaba47±2local~12GB256kLM Studio › Qwen (Alibaba) vendor-reported benchmark scores for its own models (qwen.ai launch posts, Qwen2.5-VL and Qwen3-VL technical reports, Qwen3.8-Flash-Next card)ChartArena: Benchmarking Chart Parsing across Languages, Scenarios, and Formats (arXiv:2606.01348v3)
119Qwen3-VL 30B-A3B Instruct · 4-bit Alibaba46±2†local~21GB256kllama.cpp › Qwen (Alibaba) vendor-reported benchmark scores for its own models (qwen.ai launch posts, Qwen2.5-VL and Qwen3-VL technical reports, Qwen3.8-Flash-Next card)
120Magistral Small 1.2 · 4-bit Mistral46±5*local~17GB131kllama.cpp › Apriel-1.5-15b-Thinker (arXiv:2510.01141v1), Table 4 multimodal benchmarks
125Mistral Small 3.2 24B · 4-bit Mistral44±2local~17GB128kOllama › LMArena leaderboard datasetVMMU: A Vietnamese Multitask Multimodal Understanding and Reasoning Benchmark (arXiv:2508.13680v5)mistralai/Mistral-Small-3.2-24B-Instruct-2506 model cardmistralai/Mistral-Small-4-119B-2603 model card
126Qwen3-VL 4B Thinking · 8-bit Alibaba44±2local~6.9GB256kOllama › Qwen (Alibaba) vendor-reported benchmark scores for its own models (qwen.ai launch posts, Qwen2.5-VL and Qwen3-VL technical reports, Qwen3.8-Flash-Next card)ChartArena: Benchmarking Chart Parsing across Languages, Scenarios, and Formats (arXiv:2606.01348v3)BEAR-Bench: A Bilingual Enterprise and Academic Reasoning Benchmark for Multimodal Models (arXiv:2608.17895v1)
129Qwen3-VL 8B Instruct · 4-bit Alibaba44±2local~12GB256kLM Studio › Qwen (Alibaba) vendor-reported benchmark scores for its own models (qwen.ai launch posts, Qwen2.5-VL and Qwen3-VL technical reports, Qwen3.8-Flash-Next card)ChartArena: Benchmarking Chart Parsing across Languages, Scenarios, and Formats (arXiv:2606.01348v3)BEAR-Bench: A Bilingual Enterprise and Academic Reasoning Benchmark for Multimodal Models (arXiv:2608.17895v1)How Far Is Document Parsing from Solved? PureDocBench: A Source-Traceable Benchmark across Clean, Degraded, and Real-World Settings (arXiv:2605.07492v2)
131Mistral Small 3.1 24B (2503) · 4-bit Mistral44±2local~17GB128kLM Studio › LMArena leaderboard datasetMistral AI vendor-reported benchmark scores for its own models (Hugging Face model cards)mistralai/Mistral-Small-3.2-24B-Instruct-2506 model card
133Gemma 3 27B · 4-bit Google42±2local~19GB131kOllama › LMArena leaderboard datasetGemma Team — Gemma 4 Technical ReportGoogle DeepMind Gemma 3 270M IT model cardVisEditBench: Can Vision-Language Models Edit Visualization Code from Multimodal Feedback? (arXiv:2608.10408v1)VMMU: A Vietnamese Multitask Multimodal Understanding and Reasoning Benchmark (arXiv:2508.13680v5)Gemma Team — Gemma 3 Technical Report
135Molmo 2 8B · 8-bit Ai242±2local~11GBllama.cpp › LMArena leaderboard datasetMolmo2: Open Weights and Data for Vision-Language Models with Video Understanding and Grounding (Ai2, arXiv 2601.10611v4)
138Gemma 4 E4B · 4-bit Google40±4†local~11GB131kLM Studio › Gemma Team — Gemma 4 Technical Report
140Qwen3-VL 4B Instruct · 8-bit Alibaba40±2local~6.9GB256kOllama › Qwen (Alibaba) vendor-reported benchmark scores for its own models (qwen.ai launch posts, Qwen2.5-VL and Qwen3-VL technical reports, Qwen3.8-Flash-Next card)ChartArena: Benchmarking Chart Parsing across Languages, Scenarios, and Formats (arXiv:2606.01348v3)How Far Is Document Parsing from Solved? PureDocBench: A Source-Traceable Benchmark across Clean, Degraded, and Real-World Settings (arXiv:2605.07492v2)
141Qwen3.5 2B · 8-bit Alibaba39±2local~4.9GB262kOllama › Qwen/Qwen3.5-2B model cardCohereLabs/North-Micro-Vision-Instruct model card (benchmark table)ChartArena: Benchmarking Chart Parsing across Languages, Scenarios, and Formats (arXiv:2606.01348v3)SpaceCast-Bench: Evaluating Predictive Spatial Reasoning in Vision-Language Models (arXiv:2610.12402v1)How Far Is Document Parsing from Solved? PureDocBench: A Source-Traceable Benchmark across Clean, Degraded, and Real-World Settings (arXiv:2605.07492v2)
148Qwen2.5-VL 7B Instruct · 8-bit Alibaba35±2local~11GBOllama › Qwen (Alibaba) vendor-reported benchmark scores for its own models (qwen.ai launch posts, Qwen2.5-VL and Qwen3-VL technical reports, Qwen3.8-Flash-Next card)ChartArena: Benchmarking Chart Parsing across Languages, Scenarios, and Formats (arXiv:2606.01348v3)
150Qwen3-VL 2B Thinking · 8-bit Alibaba34±2local~4.2GB256kOllama › Qwen (Alibaba) vendor-reported benchmark scores for its own models (qwen.ai launch posts, Qwen2.5-VL and Qwen3-VL technical reports, Qwen3.8-Flash-Next card)ChartArena: Benchmarking Chart Parsing across Languages, Scenarios, and Formats (arXiv:2606.01348v3)
155Pixtral 12B (2409) · 8-bit Mistral30±2local~15GBllama.cpp › LMArena leaderboard datasetPixtral 12B (Mistral AI, arXiv 2410.07073v2)VisEditBench: Can Vision-Language Models Edit Visualization Code from Multimodal Feedback? (arXiv:2608.10408v1)Ai2 Molmo-72B-0924 model card: Human Preference Elo Rating column
156Gemma 3 12B · 8-bit Google30±4local~17GB131kOllama › Google DeepMind Gemma 3 270M IT model cardVisEditBench: Can Vision-Language Models Edit Visualization Code from Multimodal Feedback? (arXiv:2608.10408v1)Gemma Team — Gemma 3 Technical Report
161Qwen2-VL 7B Instruct · 8-bit Alibaba29±1local~11GBllama.cpp › Qwen (Alibaba) vendor-reported benchmark scores for its own models (qwen.ai launch posts, Qwen2.5-VL and Qwen3-VL technical reports, Qwen3.8-Flash-Next card)LMArena leaderboard datasetPixtral 12B (Mistral AI, arXiv 2410.07073v2)CapArena: Benchmarking and Analyzing Detailed Image Captioning in the LLM Era (arXiv:2503.12329v1)Ai2 Molmo-72B-0924 model card: Human Preference Elo Rating column
163Ministral 3 3B · 8-bit Mistral27±4local~6.9GB262kOllama › CohereLabs/North-Micro-Vision-Instruct model card (benchmark table)
164Qwen3-VL 2B Instruct · 8-bit Alibaba27±1local~4.2GB256kOllama › Qwen (Alibaba) vendor-reported benchmark scores for its own models (qwen.ai launch posts, Qwen2.5-VL and Qwen3-VL technical reports, Qwen3.8-Flash-Next card)CohereLabs/North-Micro-Vision-Instruct model card (benchmark table)ChartArena: Benchmarking Chart Parsing across Languages, Scenarios, and Formats (arXiv:2606.01348v3)BEAR-Bench: A Bilingual Enterprise and Academic Reasoning Benchmark for Multimodal Models (arXiv:2608.17895v1)How Far Is Document Parsing from Solved? PureDocBench: A Source-Traceable Benchmark across Clean, Degraded, and Real-World Settings (arXiv:2605.07492v2)
165Gemma 4 E2B · 4-bit Google27±3local~7.7GB131kLM Studio › CohereLabs/North-Micro-Vision-Instruct model card (benchmark table)Gemma Team — Gemma 4 Technical Report
166Qwen2.5-VL 3B Instruct · 8-bit Alibaba27±2local~6.3GBOllama › Qwen (Alibaba) vendor-reported benchmark scores for its own models (qwen.ai launch posts, Qwen2.5-VL and Qwen3-VL technical reports, Qwen3.8-Flash-Next card)ChartArena: Benchmarking Chart Parsing across Languages, Scenarios, and Formats (arXiv:2606.01348v3)
167Qwen3.5 0.8B · 8-bit Alibaba26±2local~3.1GB262kOllama › Qwen/Qwen3.5-2B model cardChartArena: Benchmarking Chart Parsing across Languages, Scenarios, and Formats (arXiv:2606.01348v3)BEAR-Bench: A Bilingual Enterprise and Academic Reasoning Benchmark for Multimodal Models (arXiv:2608.17895v1)How Far Is Document Parsing from Solved? PureDocBench: A Source-Traceable Benchmark across Clean, Degraded, and Real-World Settings (arXiv:2605.07492v2)
170North Micro Vision Instruct (2.4B) · 4-bit Cohere24±4†local~3.8GB128khow to run › CohereLabs/North-Micro-Vision-Instruct model card (benchmark table)
173Llama 3.2 11B Vision Instruct · 8-bit Meta22±2local~16GB128kOllama › LMArena leaderboard datasetPixtral 12B (Mistral AI, arXiv 2410.07073v2)
176MiniCPM-V 2.6 · 8-bit OpenBMB21±2local~12GBOllama › LMArena leaderboard datasetCharXiv leaderboard (princeton-nlp/CharXiv README)CapArena: Benchmarking and Analyzing Detailed Image Captioning in the LLM Era (arXiv:2503.12329v1)PhysBench: Benchmarking and Enhancing Vision-Language Models for Physical World Understanding (arXiv:2501.16411v2)
180Qwen2-VL 2B Instruct · 4-bit Alibaba14±2†local~3.9GB33kLM Studio › Qwen (Alibaba) vendor-reported benchmark scores for its own models (qwen.ai launch posts, Qwen2.5-VL and Qwen3-VL technical reports, Qwen3.8-Flash-Next card)
181Gemma 3 4B · 8-bit Google14±4local~7.4GB131kOllama › Google DeepMind Gemma 3 270M IT model cardVisEditBench: Can Vision-Language Models Edit Visualization Code from Multimodal Feedback? (arXiv:2608.10408v1)Gemma Team — Gemma 3 Technical ReportVMMU: A Vietnamese Multitask Multimodal Understanding and Reasoning Benchmark (arXiv:2508.13680v5)

+ 146 more don't fit your machine or budget · show them

+ 1 models with fewer than 3 results left off, since one or two tests say little about overall ability · show them

184 of 202 catalogue models listed here · sources and licences

* new or thinly tested: ranked on too few benchmark results · † only the maker’s own numbers so far · ~RAM is estimated; full context may need more memory