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

Best AI models for web search to run locally on a 16GB Mac

A leaderboard of the best AI models for web search you can run locally on a Mac with 16GB of memory. Models are ranked on web research tests like BrowseComp, with search turned on. Each model shows the download that fits and the command to run it.

14 models · scores combine 12 benchmarks · updated 2026-10-11 02:20 UTChow it works · find a leaderboard · benchmark your machine

task[all tasks][agent tasks][coding][chat & writing][math][vision & documents][web search]

Scores use reviewed web search evaluations only. Compare scores within this task. Models without task measurements are excluded.

Models for 16GB RAM · Apple Silicon → ~11GB fast · local only · reset

+ more

chip[M1][M1 Pro][M2][M2 Pro][M3][M4][M5][M6 (16GB)]

set[all sets][Arena search systems (style controlled) · 96][BrowseComp-ZH · 13][BrowseComp-ZH · 5][BrowseComp-ZH · 3][Humanity's Last Exam with tools · 3][BrowseComp with context management · 3][BrowseComp with context management · 2][Humanity's Last Exam with tools · 2][Humanity's Last Exam with tools · 2][BrowseComp · 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

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…

#	model	task score↓	where↓	speed↓	RAM↓	context↓	get it	sources

#	model	task score↓	where↓	RAM↓	context↓	get it
—	Gemma 4 12B · q4_k_m Google	not yet calibrated	local	~9.5GB	262k	Ollama ›
—	Gemma 4 E2B · mlx-8bit Google	not yet calibrated	local	~7.7GB	131k	LM Studio ›
—	Gemma 4 E4B · Q8_0 Google	not yet calibrated	local	~9.9GB	131k	llama.cpp ›
—	Granite 4.2 3B · Q8_0 IBM	not yet calibrated	local	~5.6GB	131k	llama.cpp ›
—	K2 Horizon 7B · gguf-q4_k_m Institute of Foundation Models (MBZUAI)	not scored yet	local	~7.7GB	524k	llama.cpp ›
—	LFM2.5 2.6B · mlx-8bit Liquid AI	not yet calibrated	local	~4.5GB	131k	LM Studio ›
—	LFM2.5 8B-A1B · q4_k_m Liquid AI	not yet calibrated	local	~7.4GB	128k	Ollama ›
—	MiniCPM5 2B · mlx-8bit OpenBMB	not yet calibrated	local	~4.3GB	131k	LM Studio ›
—	Nemotron 3 Nano 4B · q8_0 NVIDIA	not yet calibrated	local	~7GB	262k	Ollama ›
—	Qwen3.5 2B · q8_0 Alibaba	not yet calibrated	local	~4.9GB	262k	Ollama ›
—	Qwen3.5 4B · mlx-8bit Alibaba	not yet calibrated	local	~6.9GB	262k	LM Studio ›
—	Qwen3.5 9B · q4_k_m Alibaba	not yet calibrated	local	~8.1GB	262k	Ollama ›
—	Spark-X2.5 1.7B · bf16 iFLYTEK (SparkLLM)	not yet calibrated	local	~5.1GB	1M	Ollama ›
—	Spark-X2.5 4B · bf16 iFLYTEK (SparkLLM)	not yet calibrated	local	~10GB	1M	Ollama ›

+ 117 more don't fit your machine or budget · show them

131 of 135 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