Local Models Library

Second-generation Devstral for agentic coding, built for tool use, multi-file editing, and software engineering agents with vision support.

Integrates breakthroughs in multimodal learning, architectural efficiency, and reinforcement learning scale.

An 80B MoE model with 3B active parameters, designed for coding agents and complex tool usage.

A high-sparsity Mixture-of-Experts 80B model utilizing a hybrid attention architecture.

A 230B MoE model built specifically for coding and agentic workflows.

A vision-language model featuring upgrades to visual perception, spatial reasoning, and image understanding.

The latest version of the Qwen3 family, including both 'thinking' and 'non-thinking' variants.

A generative AI model optimized for use in everyday devices like laptops and phones.

The #1 model on Artificial Analysis Intelligence Index with a score of 50. Trained entirely on Huawei Ascend 910B chips — zero US hardware dependency. Achieves SWE-bench 77.8% and leads on multi-turn long-context tasks.

The most-used open-weight model on OpenRouter. 256K native context with interleaved thinking. Runs at 60+ tokens/sec on Apple M5 Max 128GB unified memory — the top Mac inference pick.

Community top pick for a single 24GB GPU. Supports 201 languages natively with vision-language capability built in. IFBench 76.5 and AIME 91.3 — outperforms GPT-5.2 on instruction following.

Surpasses GPT-5 on AIME 2025 and HMMT competition math. LiveCodeBench 90%. The dominant open-weight model for advanced math and competitive programming with MIT-licensed weights.

Holds the record for largest context window of any open-weight model at 10 million tokens — enough for an entire codebase. MoE architecture means only 17B parameters are active per forward pass.

Powers Cursor Composer 2 — the industry benchmark for agentic coding. #2 on the Artificial Analysis Intelligence Index. 1T total parameters with 32B active, delivering near-frontier reasoning at extreme scale.

OpenAI's first open-weight release since GPT-2. Matches o4-mini on AIME and MMLU despite being open. Proves that open-weight frontier quality is achievable — a landmark for the local AI community.

Hybrid thinking model with toggleable chain-of-thought. MATH benchmark 96.6 and AIME 92.1. The most capable local reasoning model for a dual 40GB GPU setup — beats all previous open reasoning models.

Best bilingual Chinese/English 72B model available. MATH benchmark 90.1 and CEval 92.1. 262K context window makes it the top choice for legal, academic, and enterprise document processing across languages.

Beats GPT-5.2 on IFBench (76.5 vs 75.4) and scores 72.2 on BFCL-V4 tool use. A massive MoE that activates only 17B parameters per token for high efficiency.

Winner of the 24GB VRAM agentic coding challenge. A 355B MoE model that fits a single consumer GPU by activating only 30B parameters. LiveCodeBench 89% — the most capable coding model per dollar of VRAM.

SWE-bench-lite 61.2% — one of the highest scores for an open model on real GitHub issue resolution. Built-in native diff generation and test writing. Ideal for fully automated PR creation and code review pipelines.

Runs on iPhone in airplane mode. First sub-3B model with native multimodal support — a landmark for on-device AI. Perfect for privacy-preserving mobile apps that need real conversational capability without a server.

Non-Transformer architecture with linear context scaling — never degrades on long sequences. Achieves 40,400 tokens/sec on Apple Silicon. The fastest local model for structured extraction pipelines and IoT edge nodes.

The best photorealism and text-in-image accuracy of any local model in 2026. Multi-reference image conditioning. Handles 1000-character prompts with full semantic fidelity — the definitive standard for professional AI photography.

Leading open video model for 2026. 720P with full camera motion controls and best-in-class semantic consistency across frames. The top choice for filmmakers and content creators running a single 24GB GPU workstation.

Cinematic-quality video generation accessible at just 14GB VRAM with model offloading. The most democratized cinematic video tool available. Produces Hollywood-grade motion blur, depth of field, and lighting consistency.

Matches or beats ElevenLabs Flash on prosody naturalness. 3-second voice cloning from a reference clip with no fine-tuning required. The top locally-run TTS for podcast production, audiobooks, and voice-over automation.

Best local music model for 2026. Generates up to 10 minutes of audio with precise genre, instrument, tempo, and lyrics control. Apache 2.0 commercial license. The definitive tool for indie game composers and content creators.

Top-ranked self-hosted embedding on MTEB English — outperforms all sub-72B models. 32K context window for ultra-long document encoding. The upgrade path from BGE-M3 for teams needing maximum retrieval precision.

Google's smallest Gemma 3. Runs on virtually any GPU or even CPU — great for on-device applications.

Smallest reasoning model you can run locally. Surprising chain-of-thought performance for its size.

Google's strong 4B model with multimodal capability and 128K context. One of the best small models.

Distilled reasoning power in a 7B package. Excels at math, logic, and step-by-step problem solving.

Llama-3 distilled reasoning model. Outperforms GPT-4o on several math benchmarks at 8B scale.

Google's mid-tier Gemma 3. Multimodal capable, 128K context, strong multilingual reasoning.

Qwen-2.5 distilled reasoning model. Strong chain-of-thought and math at an accessible VRAM cost.

Highly capable 22B model from Mistral AI. Excellent instruction following for enterprise chat applications.

Google's largest open Gemma model. Competes with 70B-class models from previous generations.

The sweet spot for local reasoning. Competitive with o1-mini on math and coding tasks at 32B scale.

Alibaba's advanced reasoning model. Extended thinking and reflection enables GPT-o1-level problem solving.

Full-scale distilled reasoning model. Matches OpenAI o1 on math olympiad and competitive programming tasks.

Open source coding models specializing in coding and tool calling, based on a new base model.

A lightweight, open model built as a foundation for creating specialized function calling models.

A general-purpose reasoning and chat model trained from scratch by NVIDIA, featuring a low-latency MoE architecture.

A 9B vision-language model optimized for local deployment and low-latency applications.

A family of open-weight, dense models trained from scratch by Essential AI.

A highly cost-effective, high-performing 8B instruction tuned model.

A large-scale 120B parameter model designed for high-performance reasoning.

A large hybrid model family designed specifically for efficient on-device deployment.

A family of open language models designed to enable scientific research into language modeling.

A specialized Vision Language Model (VLM) for optical character recognition tasks.

Lightweight open models supporting multilingual tasks, RAG, coding, and tool use.

An advanced reasoning model with flexible 'thinking budget' control and self-reflection capabilities.

OpenAI's open-weight LLM, supporting configurable reasoning efforts (low, medium, high).

A coding model from Mistral AI designed for codebase exploration and engineering agents.

An open-weight reasoning model capable of long chains of thought before answering.

A lightweight open model focused on high-quality, reasoning-dense synthetic data.

Microsoft's best edge release. Fits 8GB RAM and runs fast on M1 MacBook Air in airplane mode. Exceptional at structured reasoning for its size — the top choice for on-device personal assistants and document Q&A.

Math Index 91.0 — the highest math score at the 8GB VRAM tier. NVIDIA's distilled Llama-3.1 with proprietary reward model training. Ideal for STEM tutoring and quantitative analysis on a single mid-range GPU.

High-quality embedding model with 8K context. Outperforms OpenAI text-embedding-ada-002 on MTEB benchmark.

State-of-the-art embedding model for retrieval tasks. Ranks #1 on multiple MTEB categories.

Smallest Qwen 2.5 — blazing fast on any hardware. Surprisingly capable for its size on simple tasks.

Compact Llama 3.2 with impressively long 128K context window. Perfect for edge deployment.

Excellent multilingual capabilities for its size. Particularly strong in Chinese and coding tasks.

Tiny but capable vision language model. Describe images, read text, answer visual questions locally.

Best-in-class 3B model with 128K context. Outperforms many 7B models on common benchmarks.

Compact and fast. Excellent multilingual and instruction-following performance at tiny VRAM cost.

Microsoft's multilingual tiny model with enormous 128K context window. Exceptional for document tasks.

Compact code-specialized model. Strong at code completion and debugging on very limited hardware.

Highly competitive 7B model with long context and strong multilingual support. A top value pick.

Best-in-class 7B code model. Excellent at multi-language completion, bug fixing, and code explanation.

Google's code-tuned Gemma variant. Excellent at code completion tasks inside IDEs.

Meta's powerhouse 8B model with 128K context. Excellent all-rounder for chat, code, and reasoning.

State-of-the-art open-weight vision model. Analyze charts, read documents, describe complex scenes.

128K context window in a 12B model. Joint Mistral AI & NVIDIA collaboration — excellent for long-document tasks.

Exceptional 14B all-rounder. Competitive with many 30B+ models on reasoning and coding benchmarks.

Top-tier code model for 12GB+ GPUs. Strong at agentic coding, multi-file edits, and complex refactors.

Microsoft's flagship small model. Trained on synthetic data with exceptional reasoning and STEM performance.

Specialized code completion model trained on 600+ programming languages. Top-tier for in-IDE completions.

Mistral's dedicated code model. Industry-leading performance on FIM (fill-in-middle) and complex code generation.

Near-frontier performance from a local 32B model. Exceptional multilingual reasoning and instruction following.

The best open-weight code model available. Matches GPT-4o on coding benchmarks, runs locally at 20GB VRAM.

Meta's flagship 70B. Competitive with GPT-4o on many tasks. Requires dual-GPU or Mac Studio for local use.

The original instruction-tuned 70B Llama 3.1. Multi-lingual, long-context, and highly capable.

Qwen's largest open model. Exceptional multilingual and coding capabilities — among the best open-weight 70B+ models.

Meta's best vision model. Near-frontier document understanding and complex scene analysis. Needs multi-GPU or Mac Studio Ultra.

Gold standard MoE for performance-per-VRAM on a 48GB system. Battle-tested JSON mode and function calling. The preferred baseline for enterprise RAG pipelines needing 65K context without multi-GPU complexity.

HumanEval 92% — the best single-GPU coding model for RTX 5090 owners. Excels at multi-file edits, refactoring, and agentic tool use pipelines. The definitive open coding model at the 20GB VRAM tier.

The 'OLMo of coding' — fully transparent training pipeline with HumanEval 83.5%. Every component is open: weights, data, and methodology. The most trustworthy small coding model for compliance-sensitive teams.

HumanEval 76.8% at just 6.7B — beats models 10× its size through OSS-Instruct training on real open-source code. The best option for code completion on 6GB VRAM with quality that defies the parameter count.

Runs in-browser via WebGPU — no installation required. Best for Electron apps and Raspberry Pi deployments. HuggingFace's most downloaded edge model with an Apache 2.0 license and full community model ecosystem.

4-step generation with Apache 2.0 commercial license. The fastest high-quality local image model — produces studio-grade output in under 3 seconds on a 24GB GPU. The go-to for commercial product photography pipelines.

50-step SDXL quality in just 4 steps via adversarial diffusion distillation. Fully Apache 2.0 commercial. ComfyUI-native workflow. The fastest path from creative brief to production asset on 8GB VRAM hardware.

Best text-in-image rendering and the largest open LoRA fine-tune ecosystem with 50,000+ community models. The creative industry's preferred foundation for style transfer, brand consistency, and character sheet generation.

6-second 720P clips on a single 16GB GPU — the easiest entry point into local video generation. Apache 2.0 commercial license with active ComfyUI integration. Best for social media automation and rapid prototyping.

CPU-only TTS that runs on Raspberry Pi. The best quality-per-watt ratio of any voice model — 82M parameters producing studio-quality speech synthesis. Apache 2.0 commercial license with no GPU requirement whatsoever.

Flow-matching TTS with no duration modeling — produces the most natural prosody and sentence rhythm of any local voice model. MIT license. The preferred choice for giving local AI agents a human-sounding voice interface.

The de facto standard for local speech-to-text. Word-level timestamps, speaker diarization, and 99 language support. Essential for transcription pipelines, meeting summarization, and building voice-first AI interfaces.

Video + multi-image + text understanding at 8B parameters. The best vision model for 8GB VRAM setups — handles 40-frame video clips, multi-image comparison, and document understanding in a single context window.

Captioning, object detection, grounding, OCR, and segmentation in one 770M model — MIT license. The Swiss Army knife of computer vision. Runs on almost any GPU and powers automated image tagging pipelines at scale.

Click anywhere on an image or video → instant object segmentation. Apache 2.0. Universal segmentation model used in medical imaging, autonomous driving datasets, and content creation. Zero training required for any object class.

The default local RAG embedding model for 2026. 100 languages, 8K context, and three retrieval modes (dense, sparse, multi-vector) in one model. MIT license. Used in production by thousands of enterprise RAG pipelines.

The model that proved smaller can beat bigger. Mistral 7B outperforms many 13B models with blazing fast speed.

The classic vision-language model. Describe images, answer visual questions locally. Proven and reliable.

Mixture-of-Experts: 46.7B total params but only 12.9B active per token. Near-70B quality at lower inference cost.

High-quality vision model at 34B scale. Significantly better image analysis than the 7B version.

Ultra-compact sentence embedding model. Perfect for semantic search and RAG pipelines on any hardware.

Outperforms Scout on reasoning, instruction following, and mathematical capabilities while offering 1M context length. The flagship enterprise model from Meta.

Topped the LMSYS Arena leaderboard for prolonged periods. Exceptional Chinese/English bilingual fluency and high accuracy on long-document summarization tasks.

Outstanding performance at the 24GB VRAM tier with a native Google vision-language encoder integrated directly into the weights. MMLU-Pro score of 67.5.

Open safety models built on the gpt-oss foundation to help classify and filter text content.

A medium-sized Mixture-of-Experts foundation model from Baidu.

The most transparent large language model ever released — weights, training data, code, evaluation harnesses, and training logs all fully public. Essential for research reproducibility and academic citation.

Top multilingual and European benchmark performer. Unrivaled for local RAG implementations that need to parse documents natively across 23+ global languages.

The UAE TII flagship dense model. Currently one of the only truly permissive, royalty-free commercial licenses available at the 150B+ parameter scale.

The industry standard for strict JSON output and multi-step agentic workflows. By far the most reliable open model when interacting directly with APIs and backend schemas.

Native support for 23 languages. The single best open multilingual foundation model available for researchers building localized or translation-heavy pipelines.

A HuggingFace DPO alignment fine-tune built on top of Mixtral 8x22B. Renowned for its warm, helpful tone and high consumer chat preference win rates.

Tiny model with 4K native output capability — best image-per-VRAM-dollar ratio available. Apache 2.0. Runs on GTX 1070 8GB. The only model sub-1GB that produces print-resolution imagery with coherent composition.

Full open-source Sora replication — weights, training code, and full data pipeline are all public. The most transparent video generation model. Essential for researchers studying text-to-video architectures and training dynamics.

Photorealistic lip sync at 30+ FPS in real time. The best model for live avatar streaming and talking-head video creation. MIT license with active Discord community. Integrates natively with OBS and streaming tools.

Encodes PDF pages as images — bypasses all broken PDF parsers for perfect scanned document retrieval. Apache 2.0. The breakthrough for RAG on government forms, research papers, and historical archives with graphical content.

Single image → metrically accurate 3D depth map in under 0.3 seconds. Free for research use. Powers 3D scene reconstruction, bokeh simulation, and AR/VR depth estimation pipelines without any calibration data.

One photo → talking head video with natural head movement, blinks, and mouth articulation. MIT license. No video sample required — just a single still image and an audio clip. The most accessible local avatar creation tool.

Meta's flagship 4B music model with melody conditioning from reference audio clips. Best for cinematic scoring and mood-driven generation. CC-BY-NC-4.0. Industry standard for AI-assisted film and game soundtrack production.