Local AI Models

Balanced 27B model with strong reasoning. Runs on 16GB VRAM with Q4 quantization.

ollama pull qwen3.5:27b

Ultra-compact 3B model for edge devices and low VRAM setups. Runs on 4GB VRAM.

ollama pull qwen3.5:3b

Alibaba's flagship 72B model with exceptional multilingual capabilities and strong reasoning. Requires multi-GPU or high VRAM setup.

ollama pull qwen3.5:72b

Highly capable 9B model, excellent for consumer hardware. Punches well above its weight class in reasoning tasks.

ollama pull qwen3.5:9b

Alibaba's 27B dense multimodal model. SWE-bench Verified 77.2 — surpasses Qwen 3.5 397B on coding. Native image/video/text input; compatible with Claude Code and Qwen Code tooling.

ollama pull qwen3.6:27b

Alibaba's sparse MoE with 35B total and only 3B active params. Major jump on coding benchmarks vs Qwen 3.5, with inference cost closer to a 3B dense model. Abliterated variant also on HuggingFace.

ollama pull qwen3.6:35b-a3b

Ant Group's 1T MoE with 50B active params and 262K context. AA Intelligence Index #2/40 (score 34 vs mean 13). $0.30/$2.50 per M tokens; Flash variant on OpenRouter. Open weights upcoming.

DeepSeek's 7B reasoning-focused distilled model. Strong chain-of-thought reasoning, runs on 8GB VRAM.

ollama pull deepseek-r1:7b

Distilled Flash variant of V4 Pro. Near-Pro performance on simple agent tasks at 1/12 the cost ($0.14/$0.28 per M tokens). Inference gap closes to zero on routine tasks vs V4 Pro. Not local-runnable on consumer hardware.

DeepSeek's frontier 1.6T MoE, trained with FP4 QAT at scale. Novel Hybrid Attention (CSA+HCA) and Manifold-Constrained Hyper-Connections replace residuals. $1.74/$3.48 per M tokens via API; not local-runnable.

Google's 27B MoE model with only 4B active parameters per token. Near-frontier quality at a fraction of compute cost.

ollama pull gemma4:27b

Google's flagship dense 31B model with 256K context. Near-frontier quality, top open-source performer on code and reasoning. Arena Elo ~1452.

ollama pull gemma4:31b

Google's ultra-compact multimodal MoE. Only 2.3B active params with full text/image/audio support. Lowest VRAM entry point in the Gemma 4 family.

ollama pull gemma4:e2b

Google's efficient 4B-active MoE model. Excellent performance per compute unit, runs on modest consumer hardware.

ollama pull gemma4:e4b

Gemma 4 MoE variant with 4B active params from a 26B pool. Unsloth achieved best-in-class GGUF across 22 quantization levels — broadest quant coverage in the Gemma 4 family.

ollama pull gemma4:26b-a4b

Qwen 3.5 35B-A3B base with 35 domain LoRA experts and an automatic router. Embedded engineering specialist with Aeon long-term memory. Built on FineFab; fully open-source LoRA routing reference.

Meta's efficient MoE model with an unprecedented 10M token context window. 17B active parameters from a 109B total pool.

ollama pull llama4:scout

Microsoft's 14B model with exceptional reasoning for its size. Particularly strong on math, science, and STEM tasks.

ollama pull phi4

MiniMax's self-evolving MoE model with 1M token context. Recently open-sourced under Apache 2.0 license.

ollama pull minimax-m2.7

Mistral's efficient 22B model with strong instruction following and multilingual support. Apache 2.0 license.

ollama pull mistral-small

Moonshot's flagship 1T MoE with 32B active params and 256K context. BenchLM 89.9 (#6/111); SWE-bench on par with Claude Opus 4.6; supports 300 parallel sub-agents. Needs 584GB+ RAM+VRAM for Q4 local — cloud/API recommended.

Hybrid Mamba-2 + MoE + Attention architecture. Native unified four-modality (text/image/video/audio) understanding; best-in-class on MMLongBench-Doc / OCRBenchV2 / VoiceBench. Mamba layers give ~4× compute efficiency. 256K context (up to 1M). BF16 on HuggingFace, Unsloth GGUF and free OpenRouter tier available.

ollama pull nemotron3-nano-omni

OpenAI's new flagship multimodal model — marketed as 'genius-grade visual IQ'. Closed weights, API-only; listed as a cloud reference point for local-model benchmarking.

GPT-5.5's deep-reasoning tier: passes Gödel-style tests and has cracked open math conjectures. API-only premium model; included strictly as a frontier comparison anchor.

Edge-class 1.2B dense model in the MiniCPM 4 line. Targets phones, edge boxes, and laptop CPUs. Competes with Qwen 3.6 sub-3B distills for on-device workloads.

ollama pull minicpm:4.6

Apache-2.0 33B A3B MoE focused on agentic coding. Roughly matches Qwen 3.5 35B-A3B on agentic benchmarks while staying fully open-weight.

ollama pull laguna-xs2

Edge-class ternary model — runs on phones and small embedded devices. 1.58-bit quantization at 1.7B parameters. MLX packed only today; llama.cpp / vLLM ports in progress.

Mid-tier ternary model. 1.58-bit weights {-1, 0, +1} give laptop-class footprint with capability that beats dense peers at the same byte budget. MLX 2-bit packed only for now.

1.58-bit ternary quantization — weights are only {-1, 0, +1}. Memory footprint ~1/9 of FP16 at the same parameter count. MLX 2-bit packed format today; other backends coming soon.

Released alongside xAI's rebrand to SpaceXAI. Closed weights, API-only; included for completeness on the cloud-frontier comparison row.

Tencent's open-source preview MoE model. +40% efficiency gain on reasoning and coding vs prior generation. 21B active params from a 295B pool. Weights on HuggingFace; local deployment at scale needs multi-GPU.

Xiaomi's first open-source flagship: 310B Sparse MoE with 15B active, full multimodal (text/image/video/audio). llama.cpp PR #22493 landed day-one. Local run feasible only with 200GB+ aggregate memory — included as a cloud/API reference for now.

Zhipu's 357B MoE. $0.6/M tokens via API; local deployment needs 8×H200 or equivalent multi-GPU with vLLM v0.19+. Not a consumer-GPU target.

Zhipu's 744B frontier MoE. $1.0/M tokens via API. Cluster-scale deployment only; expect 200GB+ VRAM even at aggressive quantization.

Zhipu's 754B flagship MoE. $1.4/M tokens via API; strong on agentic coding benchmarks. Not local-runnable on consumer hardware — included for completeness alongside GLM-5.

Compact reasoning-tuned MoE trained on AMD GPUs, optimized for intelligence density per parameter. Comfortable fit on a single 12GB card at Q4.

ollama pull zaya1:8b