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

Best AI models for chat and writing to run locally on a 16GB Mac

A leaderboard of the best AI models for chat and writing you can run locally on a Mac with 16GB of memory. Gemma 4 12B is #1 right now, followed by Spark-X2.5 4B and Nemotron-Cascade-8B. Models are ranked on human votes from LMArena and on writing and instruction-following tests. Each model shows the download that fits and the command to run it.

148 models · scores combine 19 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 chat & writing 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][LMArena (text) · 579][IFEval (aggregation unspecified) · 3][IFEval (aggregation unspecified) · 2][IFEval (aggregation unspecified) · 2][IFEval prompt-level strict · 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][$0.02][$0.1][$0.5][$2][$10][any]max per M

detecting your OS…

Score vs memory to run

     │                                             ●─Spark-X2.5 4B──────────────────────Gemma 4 12B─●
  50 ┤ · ●───────────────────────●─LFM2.5 2.6B─────∘ · · · · · ··· · · · · ∘ · · · · ∘ · · · ∘ · · ·∘· · ·∘· ·
     │   ∘                       ∘                 ∘                       ∘         ∘       ∘      ∘     ∘
  25 ┤ · ∘ · · · · · · · · · · · ∘ · · · · · · · · ∘ · · · · · ·∘· · · · · ∘ · · · · ∘ · · · · · · ·∘· · ·∘· ·
     │                                                                               ∘              ∘
     └───┬──────────────────────────────────────────────────────┬─────────────────────────────────────────┬───
        2GB                                                    5GB                                      10GB

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

#	model	task score↓	where↓	RAM↓	context↓	get it	sources
83	Gemma 4 12B · 4-bit Google	67±6*	local	~9.5GB	262k	Ollama ›
111	Spark-X2.5 4B · 16-bit iFLYTEK (SparkLLM)	64±6*†	local	~10GB	1M	Ollama ›
124	Nemotron-Cascade-8B · 8-bit NVIDIA	63±7*†	local	~11GB		llama.cpp ›
131	LFM2.5 2.6B · 4-bit Liquid AI	62±7*	local	~4.5GB	131k	LM Studio ›
146	LFM2.5 8B-A1B · 4-bit Liquid AI	60±5*	local	~7.4GB	128k	Ollama ›
154	Spark-X2.5 1.7B · 16-bit iFLYTEK (SparkLLM)	60±5*†	local	~5.1GB	1M	Ollama ›
159	Gemma 2 9B IT DPO (Princeton) · 4-bit Princeton NLP	59±2*	local	~7.6GB		llama.cpp ›
164	LFM2.5 1.2B Thinking · 8-bit Liquid AI	59±6*†	local	~3.5GB	33k	Ollama ›
169	Qwen3.5 9B · 4-bit Alibaba	58±3	local	~8.1GB	262k	Ollama ›
176	Phi-4-reasoning-plus 14B · 4-bit Microsoft	57±3†	local	~10GB		LM Studio ›
187	Qwen3.5 4B · 4-bit Alibaba	56±3	local	~6.9GB	262k	LM Studio ›
188	Nemotron 3 Nano 4B · 8-bit NVIDIA	56±6*	local	~7GB	262k	Ollama ›
189	LFM2.5 1.2B Instruct · 4-bit Liquid AI	56±6*†	local	~2.8GB	33k	LM Studio ›
193	Mellum2.1 12B-A2.5B Thinking · 4-bit JetBrains	56±8*†	local	~10GB		llama.cpp ›
198	Gemma 3 12B · 4-bit Google	55±1*	local	~9.7GB	131k	Ollama ›
202	MiniCPM5 2B · 4-bit OpenBMB	55±6*†	local	~4.3GB	131k	LM Studio ›
205	Phi-4-reasoning 14B · 4-bit Microsoft	54±3†	local	~10GB	33k	LM Studio ›
208	Qwen3 8B · 8-bit Alibaba	54±3*†	local	~11GB	33k	llama.cpp ›
209	Qwen3 4B Instruct 2507 · 8-bit Alibaba	54±2†	local	~7.1GB	262k	Ollama ›
212	Nemotron Nano 9B v2 · 4-bit NVIDIA	54±3	local	~9.1GB	131k	LM Studio ›
215	Qwen3 4B Thinking 2507 · 8-bit Alibaba	53±3*†	local	~7.1GB	262k	Ollama ›
219	Nemotron-Cascade-14B-Thinking · 4-bit NVIDIA	53±6*†	local	~10GB		LM Studio ›
220	Qwen3-VL 8B Instruct · 8-bit Alibaba	53±3*†	local	~11GB	256k	llama.cpp ›
228	Molmo 2 8B · 8-bit Ai2	53±1*	local	~11GB		llama.cpp ›
229	Qwen3-VL 8B Thinking · 8-bit Alibaba	53±3*†	local	~11GB	256k	llama.cpp ›
230	Olmo 3 7B Think · 8-bit Ai2	53±3	local	~9.7GB	66k	llama.cpp ›
234	Qwen3-VL 4B Thinking · 8-bit Alibaba	52±3*†	local	~6.9GB	256k	Ollama ›
235	Gemma 3n E4B · 8-bit Google	52±1*	local	~9.3GB	32k	llama.cpp ›
238	Qwen3-VL 4B Instruct · 8-bit Alibaba	52±3*†	local	~6.9GB	256k	Ollama ›
240	Granite 4.0 H Micro 3B · 4-bit IBM	52±5*†	local	~5.1GB	131k	LM Studio ›
253	Granite 4.1 8B · 4-bit IBM	51±1*	local	~7.1GB	131k	llama.cpp ›
254	Gemma 2 9B IT SimPO · 4-bit Princeton NLP	51±1*	local	~7.6GB		llama.cpp ›
256	Olmo 3 7B Instruct · 8-bit Ai2	50±3	local	~9.7GB	66k	llama.cpp ›
258	Gemma 3 4B · 8-bit Google	50±1	local	~7.4GB	131k	Ollama ›
259	Granite 4.1 3B · 4-bit IBM	50±5*†	local	~3.7GB		Ollama ›
261	Granite 4.0 Micro 3B · 8-bit IBM	50±5*†	local	~5.3GB	131k	llama.cpp ›
262	Command R7B (12-2024) · 8-bit Cohere	49±3*	local	~10GB	128k	llama.cpp ›
266	Llama-3-Instruct-8B-SimPO-ExPO · 8-bit Chujie Zheng (ExPO)	49±2*	local	~10GB		llama.cpp ›
267	Granite 4.2 3B · 8-bit IBM	49±1	local	~5.6GB	131k	llama.cpp ›
268	Granite 4.2 8B · 4-bit IBM	49±1*	local	~7.6GB	131k	Ollama ›
274	Llama-3-Instruct-8B-SimPO · 8-bit Princeton NLP	49±2*	local	~10GB		llama.cpp ›
275	Granite 4.0 H Tiny (7B-A1B) · 4-bit IBM	49±4*†	local	~9.3GB	131k	LM Studio ›
277	Granite 4.0 H 1B · 8-bit IBM	48±4	local	~3.2GB	131k	Ollama ›
283	Mellum2 12B-A2.5B Instruct · 4-bit JetBrains	48±4*†	local	~10GB	131k	llama.cpp ›
285	LFM2 2.6B · 4-bit Liquid AI	48±6*†	local	~4.3GB	33k	LM Studio ›
286	Qwen3 4B · 8-bit Alibaba	47±2	local	~7.1GB	33k	Ollama ›
288	Granite 4.0 1B · 16-bit IBM	47±3	local	~6.1GB	128k	Ollama ›
289	OLMo 2 13B Instruct (1124) · 4-bit Allen Institute for AI	47±3*†	local	~11GB	4k	Ollama ›
290	Qwen3.5 2B · 8-bit Alibaba	47±3	local	~4.9GB	262k	Ollama ›
291	Qwen2.5 14B Instruct · 4-bit Alibaba	47±3	local	~10GB	131k	LM Studio ›
296	Mellum2 12B-A2.5B Thinking · 4-bit JetBrains	47±3*†	local	~10GB	131k	llama.cpp ›
299	Qwen3-VL 2B Thinking · 8-bit Alibaba	47±3*†	local	~4.2GB	256k	Ollama ›
304	North Micro Vision Instruct (2.4B) · 4-bit Cohere	46±8*†	local	~3.8GB	128k	how to run ›
306	Gemma 2 9B IT · 4-bit Google	46±1	local	~7.8GB	8k	Ollama ›
309	Qwen2.5 7B Instruct · 8-bit Alibaba	45±2	local	~10GB	131k	llama.cpp ›
315	Granite 3.3 8B Instruct · 8-bit IBM	45±4*†	local	~11GB	128k	llama.cpp ›
316	LFM2 1.2B · 4-bit Liquid AI	45±4*†	local	~2.8GB	33k	LM Studio ›
319	Phi-4 14B · 4-bit Microsoft	44±1	local	~10GB	16k	LM Studio ›
321	Granite 3.2 8B Instruct · 8-bit IBM	44±4*†	local	~11GB	131k	llama.cpp ›
322	EXAONE 3.5 2.4B Instruct · 8-bit LG AI Research	44±3*†	local	~5.1GB	33k	Ollama ›
323	Qwen3 1.7B · 8-bit Alibaba	44±2	local	~4.6GB	33k	Ollama ›
324	Qwen2.5 3B Instruct · 8-bit Alibaba	44±3*	local	~5.9GB	33k	Ollama ›
325	Qwen2 7B Instruct · 8-bit Alibaba	44±3*	local	~10GB	131k	llama.cpp ›
327	Ministral 3 14B · 4-bit Mistral	43±3*	local	~10GB	262k	llama.cpp ›
328	LFM2 700M · 4-bit Liquid AI	43±4*†	local	~2.3GB	33k	LM Studio ›
330	Qwen3-VL 2B Instruct · 8-bit Alibaba	43±2*	local	~4.2GB	256k	Ollama ›
334	SmolLM3 3B · 8-bit Hugging Face	42±4*	local	~5.9GB	66k	llama.cpp ›
335	EXAONE 4.0 1.2B · 8-bit LG AI Research	42±3*†	local	~3.6GB	66k	llama.cpp ›
337	Ministral 8B (2410) · 8-bit Mistral	42±1	local	~10GB	128k	llama.cpp ›
338	Ministral 3 8B · 4-bit Mistral	42±6*†	local	~7.6GB	262k	Ollama ›
339	OLMo 2 7B Instruct (1124) · 8-bit Allen Institute for AI	42±4*†	local	~9.7GB	4k	llama.cpp ›
342	Aya Expanse 8B · 8-bit Cohere	41±1*	local	~10GB	8k	llama.cpp ›
345	Llama 3.1 Tulu 3 8B · 8-bit Ai2	40±1	local	~10GB	131k	llama.cpp ›
351	Llama 3 8B Instruct · 8-bit Meta	39±1	local	~10GB	8k	llama.cpp ›
353	LFM2 350M · 4-bit Liquid AI	39±4*†	local	~1.9GB	33k	LM Studio ›
355	Llama 3.1 8B Instruct · 8-bit Meta	39±1	local	~10GB	128k	llama.cpp ›
356	Granite 3.1 8B Instruct · 8-bit IBM	39±1	local	~11GB	128k	llama.cpp ›
359	Gemma 3 1B · 4-bit Google	38±2	local	~3GB	32k	LM Studio ›
360	Qwen3 0.6B · 8-bit Alibaba	38±3*	local	~3GB	33k	Ollama ›
361	Ministral 3 14B Reasoning 2512 · 4-bit Mistral	38±3*	local	~10GB	262k	LM Studio ›
363	Phi-3-medium-4k-instruct · 4-bit Microsoft	38±1	local	~11GB	4k	llama.cpp ›
364	Gemma 2 2B IT · 8-bit Google	37±1*	local	~5.3GB	8k	Ollama ›
371	Mistral 7B Instruct v0.3 · 8-bit Mistral	36±2*	local	~9.6GB	32k	llama.cpp ›
374	Granite 3.3 2B Instruct · 8-bit IBM	36±3*†	local	~4.3GB	128k	llama.cpp ›
376	OpenChat 3.5 0106 · 8-bit OpenChat	35±1*	local	~9.6GB	8k	llama.cpp ›
378	Granite 3.1 3B-A800M Instruct · 8-bit IBM	35±5*†	local	~5.2GB	131k	llama.cpp ›
380	Granite 3.0 8B Instruct · 8-bit IBM	35±1*	local	~11GB	4k	llama.cpp ›
381	OpenChat 3.5 · 8-bit OpenChat	35±1*	local	~9.6GB	8k	llama.cpp ›
384	Gemma 4 E4B · 8-bit Google	35±2	local	~9.9GB	131k	llama.cpp ›
385	Gemma 1.1 7B IT · 4-bit Google	35±1	local	~7.7GB	8k	LM Studio ›
386	Granite 3.1 2B Instruct · 8-bit IBM	35±1	local	~5.2GB	128k	Ollama ›
388	Starling-LM 7B beta · 4-bit Nexusflow	34±1	local	~9.7GB	8k	LM Studio ›
389	Ministral 3 3B · 8-bit Mistral	34±5*	local	~6.9GB	262k	Ollama ›
391	Granite 4.0 H 350M · 4-bit IBM	34±4*†	local	~1.9GB	33k	LM Studio ›
393	OpenHermes 2.5 Mistral 7B · 8-bit Nous Research	34±1	local	~9.6GB		llama.cpp ›
394	Granite 3.2 2B Instruct · 8-bit IBM	34±3*†	local	~5.2GB	131k	Ollama ›
395	Llama 3.2 3B Instruct · 8-bit Meta	33±1	local	~6.1GB	131k	Ollama ›
397	OLMo 2 1B Instruct (0425) · 8-bit Allen Institute for AI	33±4*†	local	~3.2GB	4k	llama.cpp ›
398	Starling-LM 7B alpha · 8-bit UC Berkeley (NEST)	33±1	local	~9.6GB	8k	llama.cpp ›
399	OpenThinker3 7B · 8-bit Open Thoughts	33±3*	local	~10GB	33k	llama.cpp ›
401	Gemma 3 270M IT · 8-bit Google	32±5*†	local	~2.4GB	32k	Ollama ›
402	Granite 3.0 2B Instruct · 8-bit IBM	32±1*	local	~5.3GB	4k	Ollama ›
403	SmolLM2 360M Instruct · 4-bit Hugging Face	32±5*†	local	~2.3GB	8k	Ollama ›
405	WizardLM 13B V1.2 · 4-bit Microsoft	31±1	local	~9.8GB		llama.cpp ›
406	Mistral 7B Instruct v0.2 · 8-bit Mistral	31±1	local	~9.6GB	32k	llama.cpp ›
407	Tulu 2 DPO 13B · 4-bit Allen Institute for AI	31±2*	local	~9.8GB		llama.cpp ›
408	Granite 3.1 1B-A400M Instruct · 8-bit IBM	31±5*†	local	~3GB	131k	Ollama ›
409	Granite 3.0 3B-A800M Instruct · 8-bit IBM	31±4*†	local	~5.3GB	4k	llama.cpp ›
410	SOLAR 10.7B Instruct v1.0 · 4-bit Upstage	31±1*	local	~8.8GB	4k	Ollama ›
411	Qwen3.5 0.8B · 8-bit Alibaba	31±4*†	local	~3.1GB	262k	Ollama ›
412	Dolphin 2.2.1 Mistral 7B · 8-bit Cognitive Computations	31±1	local	~9.6GB		llama.cpp ›
413	Qwen1.5 7B Chat · 8-bit Alibaba	31±1	local	~10GB	32k	llama.cpp ›
415	Qwen2.5 0.5B Instruct · 8-bit Alibaba	30±3*	local	~2.6GB	33k	Ollama ›
418	Phi-3-mini-4k-instruct (June 2024 update) · 4-bit Microsoft	30±1*	local	~5.8GB		LM Studio ›
419	Vicuna 13B v1.5 · 4-bit LMSYS	30±1	local	~10GB	4k	Ollama ›
420	Granite 4.0 350M · 16-bit IBM	30±4*†	local	~2.9GB	32k	Ollama ›
421	Llama 2 13B Chat · 4-bit Meta	29±1	local	~10GB	4k	Ollama ›
423	Zephyr 7B beta · 8-bit Hugging Face	29±1	local	~9.6GB		llama.cpp ›
424	Tulu 2 DPO 7B · 8-bit Allen Institute for AI	29±2*	local	~9.1GB		llama.cpp ›
425	Phi-3-mini-128k-instruct (Apr 2024) · 16-bit Microsoft	29±1	local	~9.5GB		how to run ›
426	Gemma 7B IT · 4-bit Google	29±1	local	~7.4GB	8k	Ollama ›
427	Phi-3-mini-4k-instruct (Apr 2024) · 4-bit Microsoft	29±1*	local	~9.5GB		llama.cpp ›
428	Zephyr 7B alpha · 8-bit Hugging Face	28±1	local	~9.6GB		llama.cpp ›
430	Gemma 4 E2B · 4-bit Google	28±3	local	~7.7GB	131k	LM Studio ›
431	Llama 3.2 1B Instruct · 4-bit Meta	27±1	local	~5.2GB	131k	Ollama ›	+1
434	Vicuna 7B v1.5 · 8-bit LMSYS	27±1	local	~9.1GB	4k	llama.cpp ›
435	Gemma 1.1 2B IT · 8-bit Google	27±1	local	~5.2GB	8k	Ollama ›
436	SmolLM2 1.7B Instruct · 8-bit Hugging Face	27±1	local	~4.2GB	8k	Ollama ›
437	Granite 3.0 1B-A400M Instruct · 8-bit IBM	26±4*†	local	~3GB	4k	llama.cpp ›
438	Qwen2.5 1.5B Instruct · 8-bit Alibaba	26±3	local	~4GB	33k	Ollama ›
439	Mistral 7B Instruct v0.1 · 8-bit Mistral	26±1*	local	~9.6GB	8k	llama.cpp ›
440	Llama 2 7B Chat · 8-bit Meta	25±1	local	~11GB	4k	Ollama ›
441	Ghost 7B Alpha · 8-bit Ghost X	25±2*	local	~9.6GB		llama.cpp ›
442	Nous-Hermes-13b · 4-bit Nous Research	25±2*	local	~9.8GB	2k	llama.cpp ›
443	Qwen1.5 4B Chat · 8-bit Alibaba	24±1*	local	~7GB	32k	Ollama ›
444	Gemma 2B IT · 8-bit Google	23±1	local	~5.2GB	8k	Ollama ›
451	ChatGLM3-6B · 8-bit Zhipu AI (THUDM)	19±1*	local	~8.4GB		llama.cpp ›
452	Falcon 7B Instruct · 8-bit TII	18±2*	local	~9.3GB		llama.cpp ›
455	Qwen1.5 1.8B Chat · 4-bit Alibaba	16±5*	local	~3.1GB	33k	LM Studio ›
456	Phi-2 (2.7B) · 4-bit Microsoft	14±2*	local	~7.3GB	2k	how to run ›
457	OpenAssistant SFT-4 Pythia 12B · 4-bit LAION (OpenAssistant)	14±1	local	~9.5GB		llama.cpp ›
460	FastChat-T5 3B · 4-bit LMSYS	10±1	local	~8.5GB		how to run ›
463	StableLM-Tuned-Alpha 7B · 8-bit Stability AI	6±1	local	~10GB		llama.cpp ›
—	Hermes 3 Llama 3.1 8B · Q8_0 Nous Research	not yet calibrated	local	~10GB	128k	llama.cpp ›
—	Phi-3.5-mini-instruct · q8_0 Microsoft	not yet calibrated	local	~6.8GB	128k	Ollama ›
—	Phi-4-mini 3.8B · q8_0 Microsoft	not yet calibrated	local	~6.9GB	128k	Ollama ›
—	Qwen2 0.5B Instruct · q8_0 Alibaba	not yet calibrated	local	~2.6GB	32k	Ollama ›
—	Qwen2 1.5B Instruct · q8_0 Alibaba	not yet calibrated	local	~4GB	32k	Ollama ›

+ 324 more don't fit your machine or budget · show them

+ 41 models with fewer than 3 results left off, since one or two tests say little about overall ability · show them

472 of 588 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