Model finder

Open-weight models that rank on major leaderboards, sorted by score. Pick a tier, type a name, or enter how much memory you have. Every number links to its source. Each model links to its can-it-run pages, or use What fits my memory.

Short answer: the highest-ranked open models that fit a home tier are Gemma 4 31B and 26B-A4B and Qwen3.8 27B, which need a 24–32 GB GPU. 8–16 GB GPUs run Llama 3.1 8B, Phi-4 and gpt-oss 20B. 128 GB of unified memory runs Qwen3.5 122B-A10B, gpt-oss 120B and Nemotron 3.5 Super.

Showing 13 of 13 models.

Model to hardware table
Model	LMArena text	Artificial Analysis	Memory needed	Smallest tier (our estimate)
Gemma 4 31BGoogle · 31B · Can it run?	#76 · 1452±7	15index	19–20 GBQ4_K_M · source	24–32 GB GPU24–32 GB GPU
Gemma 4 26B-A4B (MoE)Google · 26B total, 4B active · Can it run?	#100 · 1437±7	17*index	16 GBdefault tag · source	24–32 GB GPU24–32 GB GPU; 16 GB GPU only with short context
Qwen3.8 27BAlibaba Qwen · 27B · Can it run?	#102 · 1437±5	34index	18 GBQ4_K_M · source	24–32 GB GPU24–32 GB GPU
Qwen3.5 122B-A10B (MoE)Alibaba Qwen · 122B total, 10B active · Can it run?	#138 · 1417±4	16index	81 GBQ4_K_M · source	Unified 128 GBUnified 128 GB (needs most of it as GPU memory)
Nemotron 3.5 Super 120B-A12BNVIDIA · 120B total, 12B active · Can it run?	#149 · 1413±9 · prelim.	26index	90 GBQ4_K_M (default tag 80 GB) · source	Unified 128 GBUnified 128 GB (needs most of it as GPU memory)
Qwen3.5 27BAlibaba Qwen · 27B · Can it run?	#155 · 1409±4	not ranked	17–20 GBQ4_K_M · source	24–32 GB GPU24–32 GB GPU
Qwen3.5 35B-A3B (MoE)Alibaba Qwen · 35B total, 3B active · Can it run?	#172 · 1394±4	not ranked	22 GBQ4_K_M · source	24–32 GB GPU24–32 GB GPU or unified 64 GB
Mistral Small 3.2 24BMistral AI · 24B · Can it run?	#212 · 1356±5	not ranked	15 GBQ4_K_M · source	16 GB GPU16 GB GPU (tight), 24 GB comfortable
gpt-oss 120BOpenAI · 117B total, 5.1B active · Can it run?	#218 · 1352±4	12index	65 GBMXFP4 (as shipped) · source	Unified 128 GBUnified 128 GB
Llama 3.3 70BMeta · 70B · Can it run?	#266 · 1318±4	8*index	43 GBQ4_K_M · source	Unified 64 GBUnified 64 GB (tight) or 128 GB; or two 24–32 GB GPUs
gpt-oss 20BOpenAI · 21B total, 3.6B active · Can it run?	#269 · 1317±6	9index	14 GBMXFP4 (as shipped) · source	16 GB GPU16 GB GPU (tight), 24 GB comfortable
Phi-4 14BMicrosoft · 14B · Can it run?	#322 · 1256±5	6*index	9.1 GBQ4_K_M · source	8–12 GB GPU12 GB GPU (tight), 16 GB comfortable
Llama 3.1 8BMeta · 8B · Can it run?	#345 · 1211±4	not ranked	4.9 GBQ4_K_M · source	8–12 GB GPU8–12 GB GPU

The memory filter shows models whose listed size plus 2 GB of headroom fits your number. That 2 GB is a rough estimate, and long contexts need more.

Leaderboards checked Oct 27, 2026, 2:16 PM CT. LMArena Text: Overall rank among all 418 models (open and proprietary), leaderboard dated Oct 22, 2026. Score is Arena Elo ± 95% interval. Artificial Analysis Intelligence Index: Intelligence Index v4.3.2, best listed variant. * means AA marks the score with an asterisk on its leaderboard. Models that are not ranked on either board are not listed here; their guides stay up (SDXL 1.0, Whisper large / turbo, Qwen3 14B, FLUX.1 dev, DeepSeek-R1 32B (Qwen distill)).

Checked Oct 27, 2026. Ollama sizes are download sizes; add roughly 1–4 GB for context and runtime. "Tight" means it may spill into system RAM at long contexts.

FAQ

Where do the memory numbers come from?

Sizes are the Ollama library download size at Q4_K_M or the format the model ships in, checked Oct 27, 2026.

Which models are listed?

Only open-weight models that are ranked on LMArena's text leaderboard or the Artificial Analysis Intelligence Index, and that fit one of our hardware tiers. Ranks and scores are copied from those pages; we do not score models ourselves.

Why add headroom?

The runtime and the context cache use memory on top of the model file. We budget roughly 1–4 GB, more for long contexts. That headroom is our estimate.