Open model directory

43 models · leaderboard snapshot from local.ai (2026-08). Intelligence = 10% τ² + 35% GAIA + 55% GDPval. Each model page adds quant-size estimates, VRAM fit checks and the self-hosting savings calculator.

Model	Maker	Params	Intelligence	τ²	GAIA	GDPval
Step 3.7 Flash	StepFun	148B	78.8	99%	87%	70%	details →
DeepSeek V4 Flash 0731	DeepSeek	162B	77.5	85%	76%	77%	details →
DeepSeek V4 Flash	DeepSeek	162B	77.3	80%	82%	74%	details →
Qwen3.6 27B	Qwen	27B	76.5	100%	74%	74%	details →
Qwen3.6 35B A3BMoE	Qwen	35BA3B	74.9	99%	74%	71%	details →
Ornith 1.0 35B	HCompany	35B	74.1	99%	72%	71%	details →
Agents A1agents	r0b0tlab	—	72.5	93%	82%	63%	details →
Ornith 1.0 397B	HCompany	397B	72.4	97%	69%	70%	details →
Laguna S 2.1	Liquid	—	71.6	62%	73%	73%	details →
Nex N2 Mini	InternScience	—	71.3	85%	79%	64%	details →
Gemma 4 31B It	Google	31B	66.2	77%	71%	61%	details →
Nemotron 3.5 Lightning	NVIDIA	—	65.5	62%	68%	65%	details →
Nemotron 3 Super 120B A12BMoE	NVIDIA	120BA12B	64.6	68%	70%	60%	details →
Gemma 4 26B A4BMoE	Google	26BA4B	62.8	80%	65%	58%	details →
Holo 3.1 35B A3BMoE	PrismML	35BA3B	60.7	97%	59%	55%	details →
Qwen3.5 9B Base	Qwen	9B	59.3	99%	54%	55%	details →
Bonsai 27B	Xiaomi	27B	59	68%	59%	58%	details →
Laguna XS 2.1	Liquid	—	58.8	57%	64%	56%	details →
Gpt Oss 120B	OpenAI	120B	55.5	65%	56%	53%	details →
Qwen3.5 9B	Qwen	9B	55.3	98%	55%	47%	details →
Gemma 4 12B It	Google	12B	54	82%	54%	49%	details →
Diffusiongemma 26B A4B ItMoEdiffusion	Google	26BA4B	50.3	66%	45%	51%	details →
Gpt Oss 20B	OpenAI	20B	43.3	78%	44%	37%	details →
Nemotron 3 Nano 30B A3BMoE	NVIDIA	30BA3B	42.4	39%	46%	41%	details →
Mimo V2 Flash	Xiaomi	—	39	85%	38%	31%	details →
Nemotron 3 Nano Omni 30B A3B ReasoningMoEreasoning	NVIDIA	30BA3B	34.1	48%	33%	32%	details →
Gemma 4 E4B Itedge	Google	4B	28.4	14%	30%	30%	details →
Gemma 4 E2B Itedge	Google	2B	20.6	8%	26%	19%	details →
LFM2.5 8B A1BMoEedge	Liquid	8BA1B	17.9	94%	9%	10%	details →
Gemma 4 12B Coder Fable5 Composer2.5 V1	Google	—	—	—	—	—	details →
Llama 3.1 70B	Meta	70B	—	—	—	—	details →
Llama 3.1 70B Instruct	Meta	70B	—	—	—	—	details →
Llama 3.1 Nemotron 70B	NVIDIA	70B	—	—	—	—	details →
Llama 3.3 70B Instruct	Meta	70B	—	—	—	—	details →
Nemotron 3 Ultra 550B A55BMoE	NVIDIA	550BA55B	—	—	—	—	details →
Qwen2 72B Instruct	Qwen	72B	—	—	—	—	details →
Qwen2.5 32B	Qwen	32B	—	—	—	—	details →
Qwen2.5 72B	Qwen	72B	—	—	—	—	details →
Qwen2.5 72B Instruct	Qwen	72B	—	—	—	—	details →
Qwen2.5 Coder 32Bcoding	Qwen	32B	—	—	—	—	details →
Qwen3 235B A22BMoE	Qwen	235BA22B	—	—	—	—	details →
Qwen3-32B	Qwen	32B	—	—	—	—	details →
Qwen3.5 122B A10BMoE	Qwen	122BA10B	—	—	—	—	details →

Makers represented: DeepSeek, Google, HCompany, InternScience, Liquid, Meta, NVIDIA, OpenAI, PrismML, Qwen, StepFun, Xiaomi, r0b0tlab. Benchmark methodology and measurements belong to local.ai / Exo Labs; this directory is an independent presentation layer with fit and cost estimates.