Top 100 Local AI Models For Privacy + Best Outputs
A lot of us are running subscription-based AI like Claude and Codex — but as they give less for the buck and privacy concerns rise, local LLMs are the answer. No subscription fees (just electricity), fully private, and a one-time hardware investment.
In this complete 2026 guide I cover the top 100 AI models — from general purpose to coding, images, video, voice, music, and embeddings — with VRAM requirements and benchmarks so you can match every model to your hardware.
- → Will It Run? — check if your GPU can handle a specific model
- → Token Speed Estimator — predict tokens/sec before buying
- → GPU Compare — side-by-side VRAM and bandwidth
- → AI PC Builder — configure your full rig by model tier
Why Local AI in 2026
The gap between cloud and local AI has closed for most everyday tasks. GLM, Qwen, Kimi, and MiniMax show open-source is catching up fast. Privacy concerns are real — AI is increasingly a surveillance vector. VRAM and quantization improvements mean yesterday's impossible is today's default.
My rule: never buy the bare minimum. See my Budget, Mid-Range, and Elite build guides for spec recommendations. Confused about terminology? Check the full Glossary.
For a detailed cost breakdown vs. cloud subscriptions, read Local vs. Cloud Agents: The $15,000/year Cost Savings. For why VRAM headroom matters in agentic workflows, see Why AI Agents Need More VRAM.
🧠 General Purpose
GLM-5 (Z.ai)
744B / 40B active (MoE) · MIT
#1 Artificial Analysis Intelligence Index (score 50). Trained on 100K Huawei Ascend 910B chips — zero US hardware. SWE-bench 77.8%.
MiniMax M2.5
230B / 10B active (MoE) · Modified MIT
Most-used open-weight on OpenRouter. Interleaved thinking. 60+ t/s on M5 Max 128GB.
Qwen3.5-27B
27B dense · Apache 2.0
Community top pick for single 24GB GPU. 201 languages. Native vision-language. IFBench 76.5 · AIME 91.3.
Qwen3.5-397B-A17B
397B / 17B active (MoE) · Apache 2.0
Beats GPT-5.2 on IFBench (76.5 vs 75.4). BFCL-V4 tool use 72.2.
DeepSeek V3.2
685B (MoE) · DeepSeek License
Surpasses GPT-5 on AIME/HMMT 2025. LiveCodeBench 90%.
Llama 4 Scout
109B / 17B active (MoE) · Llama 4 Community
10 million token context — largest of any open-weight model.
Kimi-K2.5
1T / 32B active (MoE) · Modified MIT
#2 Intelligence Index (48). Powers Cursor Composer 2.
GPT-OSS 120B
117B (MoE) · OpenAI open
OpenAI's first open-weight model since GPT-2. Matches o4-mini on AIME/MMLU.
Llama 4 Maverick
400B / 17B active (MoE) · Llama 4 Community
Outperforms Scout on reasoning and math.
Mistral Large 3 (123B)
123B dense · Mistral Research
Top multilingual/European benchmark performer. Best for local RAG across 23+ languages.
DeepSeek R2
236B / 21B active (MoE) · MIT
Hybrid thinking: toggleable chain-of-thought. MATH 96.6 · AIME 92.1.
Yi-Lightning-2 (01.AI)
200B (MoE) · Apache 2.0
Top LMSYS Arena. Strong Chinese/English bilingual and long-doc summarization.
Falcon 3 (180B)
180B dense · TII Falcon (commercial)
UAE TII flagship. One of the only truly permissive commercial licenses at 180B.
InternLM3-72B
72B dense · Apache 2.0
Best bilingual Chinese/English 72B. MATH 90.1 · CEval 92.1.
OLMo 2 (32B)
32B dense · Apache 2.0
Allen Institute. Fully open: weights, training data, code, and evals all public.
Mixtral 8x22B
141B / 39B active (MoE) · Apache 2.0
Gold standard MoE for performance-per-VRAM. Battle-tested JSON/function calling.
Gemma 3 27B
27B dense · Gemma Terms
Native vision + text. Outstanding at the 24GB tier. MMLU-Pro 67.5.
Command A (Cohere)
111B (MoE) · CC-BY-NC-4.0
Best tool-use and JSON output accuracy of any open model. Built for multi-step agentic pipelines.
Aya Expanse 32B
32B dense · CC-BY-NC-4.0
23 simultaneous languages. Best open multilingual model for local translation.
Zephyr 141B-A39B
141B / 39B active (MoE) · Apache 2.0
HuggingFace DPO fine-tune on Mixtral 8x22B. Exceptionally helpful for consumer chat.
💻 Coding + Agentic
See also: LM Studio Guide · Ollama + Open WebUI Setup · Why Agents Need More VRAM.
GLM-4.7-Flash
355B MoE (30B active at 24GB) · MIT
Winner of the 24GB VRAM agentic coding challenge. LiveCodeBench 89% on a single consumer GPU.
Qwen3-Coder-Next (80B)
80B (MoE) · Apache 2.0
Multi-file repo editing and agentic workflows. Best Cursor-style IDE integration.
Qwen 2.5 Coder 32B
32B dense · Apache 2.0
HumanEval 92%. Best single-GPU coding model for RTX 5090 users.
Qwen 2.5 Coder 14B
14B dense · Apache 2.0
HumanEval 85%. Best coding model for 16GB VRAM. Community favorite for LM Studio.
OmniCoder-9B (Tesslate)
9B (Qwen3.5 base) · Apache 2.0
Fine-tuned on 425K agentic coding trajectories for real software engineering.
Devstral-2-123B
123B dense · Mistral Research
Best for deep refactoring, legacy codebase understanding, full-stack architecture.
DeepCoder-V2-236B
236B (MoE) · MIT
SWE-bench-lite 61.2%. Native diff generation and test writing built-in.
StarCoder 2 (15B)
15B dense · BigCode OpenRAIL-M
600+ programming languages. Gold standard for fill-in-the-middle (FIM) completion.
CodeLlama 70B
70B dense · Llama 2 Community
HumanEval 67.8%. Massive LoRA ecosystem for context-rich completion on large codebases.
SWE-Llama-3.1-70B
70B dense · Llama 3.1 Community
SWE-bench 43.8%. Specifically trained to resolve real GitHub issues end-to-end.
Jan-nano (Menlo, 4B)
4B · Apache 2.0
Plug-and-play for Jan.ai desktop. Best zero-config local code completion at 4B.
Granite 3.3 Code (34B)
34B dense · Apache 2.0
IBM. 116 programming languages. HumanEval 78.3%. Best for compliance-bound organizations.
DeepSeek R1 Distill 32B
32B dense · MIT
Self-reflection fine-tune for complex multi-step coding challenges.
OpenCoder-8B-Instruct
8B dense · MIT
Fully transparent pipeline. HumanEval 83.5%. The OLMo of coding models.
MagicCoder-S-DS-6.7B
6.7B (DeepSeek Coder) · MIT
HumanEval 76.8% — beats models 10× its size. OSS-Instruct training methodology.
📱 Edge + Mobile
Qwen3.5-9B
9B dense · Apache 2.0
Matches GPT-OSS-120B on GPQA Diamond (81.7 vs 71.5). Overperformer of 2026.
Phi-4-mini (3.8B)
3.8B dense · MIT
Fits 8GB RAM. Fast on M1 MacBook Air. Microsoft's best edge release.
Qwen3.5-2B / 4B
2B / 4B dense · Apache 2.0
Runs on iPhone in airplane mode. Native multimodal at sub-3B — unprecedented.
Nanbeige4.1-3B
3B dense · Apache 2.0
First sub-4B with 500+ round native tool invocations and deep-search.
LFM-2.5-350M (Liquid AI)
350M · Liquid AI
Non-Transformer. Linear context scaling. 40,400 t/s on Apple Silicon.
Gemma 3 4B / 12B
4B / 12B dense · Gemma Terms
Top multilingual OCR. Multimodal vision at 4B. Best for mobile RAG pipelines.
DeepSeek R1 7B Distill
7B dense · MIT
Best local reasoning for 6–8GB VRAM. MIT licensed.
Phi-4 (14B)
14B dense · MIT
MATH 80.4%. Outperforms many 70B models on structured reasoning tasks.
NVIDIA Nemotron Nano 8B
8B dense · NVIDIA
Math Index 91.0. Best math scores at the 8GB VRAM tier.
SmolLM2-1.7B
1.7B dense · Apache 2.0
Runs in-browser via WebGPU. Best for Electron apps and Raspberry Pi deployments.
MobileLLM-125M (Meta)
125M · MIT
On-device Android/iOS. No server needed. Sub-200M state-of-the-art.
H2O Danube 3 (4B)
4B dense · Apache 2.0
Document intelligence: invoices, tables, financial statements locally.
Orca Mini 3B
3B dense · CC-BY-NC-4.0
Microsoft Research synth-data. Best instruction-following at 3B. IoT proven.
🎨 Image Generation
See also: Best Local AI Image Generators Guide and Stable Diffusion XL VRAM vs Speed and Best GPU for Stable Diffusion.
FLUX.2 Dev (32B)
32B diffusion · FLUX non-commercial
Best photorealism and text-in-image accuracy. Multi-reference conditioning.
FLUX.2 Klein (4B)
4B diffusion · Apache 2.0
Real-time generation. Fully commercial Apache 2.0 license.
FLUX.1 Schnell (12B)
12B diffusion · Apache 2.0
4-step generation. Fastest high-quality local image gen. Apache 2.0 commercial.
FLUX.1 Kontext Dev
12B diffusion · FLUX non-commercial
Precise region image editing guided by reference images.
Stable Diffusion 3.5 Large
8B diffusion · Stability AI
Best text-in-image. Largest open LoRA ecosystem. Essential for creative fine-tuning.
HunyuanImage-3.0
80B MoE diffusion · Tencent
Largest image MoE. Handles 1,000-word prompts with complete semantic fidelity.
SDXL-Lightning (4-step)
~3.5B diffusion · Apache 2.0
Quality matching 50-step SDXL in 4 steps. Apache 2.0 commercial.
Kolors (Kuaishou)
~9B diffusion · Apache 2.0
Best Asian aesthetics, anime, and Chinese text-in-image rendering.
SDXL-Turbo
~3.5B diffusion · SDXL Turbo
Single-step adversarial distillation. 24fps live preview on RTX 4090.
PixArt-Σ (600M)
600M diffusion · Apache 2.0
Tiny model, 4K output. Best image-per-VRAM-dollar ratio available.
Adobe Firefly Research
~12B diffusion · Research only
Trained exclusively on licensed content. Best for legally safe creative workflows.
InstaFlow (1-step)
~1.8B · CC-BY-NC-4.0
Rectified Flow. 1-step. 100× faster than diffusion. No multi-step sampling at all.
🎬 Video Generation
See also: Best Local AI Video Generators Guide.
Wan 2.2 (Alibaba)
14B video · Apache 2.0
Leading 2026 open video model. 720P, camera motion controls, best semantic consistency.
LTX-2 (Lightricks)
~12B video · LTX License
Audio + video in one pass. 4K native output.
HunyuanVideo 1.5
~13B video · Tencent
Cinematic output. Most accessible cinematic video — runs at 14GB with offloading.
SkyReels V2
14B video · Apache 2.0
33 facial expressions, 400 natural movements. Film and TV production-grade.
CogVideoX-5B
5B video · Apache 2.0
6-second 720P clips. Best entry-point video gen on a single 16GB card.
Mochi 1 (Genmo)
~10B video · Apache 2.0
Apache 2.0 commercial. Strong photorealism and physics simulation.
Open-Sora
~4B video · Apache 2.0
Full open-source Sora replication. Weights + training code + data pipeline all public.
AnimateDiff-Lightning
~1.5B video · Apache 2.0
4-step video distillation. Near-real-time preview generation.
ModelScope T2V (1.7B)
1.7B video · CC-BY-NC-4.0
Most accessible text-to-video on a single mid-range GPU.
Pyramid Flow
~2B video · MIT
768P on single 24GB GPU. Lower compute than diffusion via autoregressive flow.
🎙️ Voice / TTS + Lip Sync
Voxtral TTS (Mistral)
~7B · Mistral
Better than ElevenLabs Flash. 3-second voice cloning from a reference clip.
Higgs Audio V2 (BosonAI)
~3B · Apache 2.0
Top trending TTS on HuggingFace. Exceptional emotional range and prosody control.
Qwen3-TTS
~3B · Apache 2.0
10 languages. Describe the voice you want in plain text. Commercial use.
Kokoro-82M
82M · Apache 2.0
CPU-only. Raspberry Pi compatible. Best quality-per-watt TTS.
Bark (Suno AI)
~300M · MIT
MIT. Laughter, sighs, background noise, and music alongside speech.
Coqui XTTS-v2
~70M · Coqui Public
17-language zero-shot voice cloning. <0.8× real-time on CPU. Best for dub pipelines.
F5-TTS
~300M · MIT
Flow-matching TTS. No duration modeling. Most natural prosody for agent voices.
Parler-TTS Mini
~880M · Apache 2.0
Describe speaker characteristics in plain text. No voice reference needed.
WhisperX
~1.5B · MIT
De facto local STT standard. Word-level timestamps and speaker diarization.
MuseTalk (Tencent)
~0.5B · MIT
Photorealistic lip sync at 30+ FPS real time. Best for live avatar streaming.
Wav2Lip
~30M · MIT
Industry standard offline film dubbing. Integrates cleanly with ComfyUI.
SadTalker
~300M · MIT
One photo → talking head video with natural movement and blinks.
LivePortrait
~200M · MIT
Emotion-aware portrait animation. Per-facial-action-unit control.
LatentSync (ByteDance)
~500M · Apache 2.0
Diffusion-based lip sync. No flickering. Best quality for post-production pipelines.
🎵 Music Generation
ACE-Step 1.5
~2B · Apache 2.0
Best 2026 local music model. Up to 10 minutes. Genre/instrument/lyrics control.
MusicGen Large (4B)
4B · CC-BY-NC-4.0
Meta. Melody conditioning from audio clips. Best for cinematic scoring.
AudioCraft Suite (Meta)
~2B · MIT
MusicGen + AudioGen + EnCodec in one package. Full local audio post-production.
👁️ Vision + Embeddings
Qwen3-VL (235B MoE)
235B MoE · Apache 2.0
Rivals Gemini 2.5 Pro. Document parsing, OCR, charts, visual reasoning.
InternVL 3.0 (108B)
108B · MIT
3D vision perception and document digitalization. Best open VLM for doc intelligence.
LLaVA-Next (34B)
34B dense · Apache 2.0
Most battle-tested local VLM. CLIP vision + Mistral text. Structured image Q&A.
MiniCPM-V 2.6 (8B)
8B · Apache 2.0
Video + multi-image + text at 8B. Best vision model for 8GB VRAM setups.
Florence-2 (Microsoft)
770M · MIT
Captioning + detection + grounding + OCR + segmentation in one tiny model.
SAM 2 (Meta)
~300M · Apache 2.0
Universal segmentation. Click anywhere → instant object segment in image or video.
Depth Pro (Apple)
~300M · Apple Sample Code
Single image → metrically accurate 3D depth map. Free for research.
BGE-M3 (568M)
568M · MIT
Default local RAG embedding. 100 languages. Dense + sparse + multi-vec retrieval.
Qwen3-Embedding-8B
8B · Apache 2.0
Top self-hosted embedding. Outperforms all sub-72B on MTEB English. 32K context.
ColPali (Vision RAG)
~3B · Apache 2.0
Encodes PDF pages as images — bypasses broken PDF parsers. Best for scanned docs.
nomic-embed-vision-v1.5
~137M · Apache 2.0
Compatible image + text vectors in one index. Zero pipeline changes for multimodal RAG.
DINO v2 (Meta)
~300M · Apache 2.0
Universal vision backbone for detection, segmentation, depth, and classification.
🖥️ Hardware Quick Reference
| VRAM | Best Models | GPU Options |
|---|---|---|
| 8GB | Phi-4-mini, Gemma 3 4B, Qwen3.5-4B, Kokoro, PixArt-Σ | RTX 4070 Super |
| 12GB | Qwen 2.5 Coder 14B, FLUX.1 Schnell, CogVideoX-5B, SDXL | RTX 3060 12GB |
| 16GB | Gemma 3 27B (Q4), HunyuanVideo 1.5, LLaVA-Next 13B | RX 9070 XT · RTX 4080 Super |
| 24GB | GLM-4.7-Flash, Qwen3.5-27B, FLUX.2 Dev, Wan 2.2 | RTX 4090 · RTX 3090 |
| 32GB+ | Qwen3.5-27B (full), Llama 4 Scout, Devstral 123B | RTX 5090 |
| Multi-GPU | GLM-5, Kimi-K2.5, DeepSeek V3.2, HunyuanImage-3.0 | RTX 5090 vs 5080 |
| Mac Unified | MiniMax M2.5, Llama 4 Scout, Mixtral 8x22B | Mac Studio M4 Max · M4 Ultra |
🔧 Tools to Run Everything
| Tool | Purpose | Link |
|---|---|---|
| Ollama | One-line LLM runner (CLI) | Website ↗ · Setup Guide |
| LM Studio | Desktop GUI for local LLMs | Website ↗ · Guide |
| Jan.ai | Offline-first desktop AI app | Website ↗ |
| ComfyUI | Node-graph for image/video gen | Website ↗ |
| Open WebUI | ChatGPT-style web interface | Website ↗ |
| llama.cpp | CPU/GPU inference engine | Website ↗ · Glossary |
Use these tools to find the right hardware for the models you want to run:
By Justin Murray · AI Computer Guide — VRAM-centric hardware validation for private, fast, local AI inference. All 100 models in this guide run fully offline, privately, with no subscription. Q2 2026.
About the Author: Justin Murray
AI Computer Guide Founder, has over a decade of AI and computer hardware experience. From leading the cryptocurrency mining hardware rush to repairing personal and commercial computer hardware, Justin has always had a passion for sharing knowledge and the cutting edge.