RunPod

NVIDIA's open multimodal mixture-of-experts model (120B total, 12B active parameters) for coding, reasoning, tool use and document understanding, with image input and a 1M-token context window.

Liquid AI's mixture-of-experts reasoning model with 24B total and 2.45B active parameters. It always reasons before answering and is built for agents that call tools.

NVIDIA's small vision-language model (8B parameters, 4B effective), post-trained from Gemma 4 E4B for fast local conversations, image understanding and multi-turn tool use; a bundled draft model speeds up generation.

Deep Reinforce's 122B MoE self-improving model (MIT) with 256K context, image input, and native tool use, the largest Ornith 1.5 size that runs on a single high-memory machine.

Prism ML's 27B vision model trained for ternary weights (-1, 0 or +1), about half the size of a standard 4-bit model at close to full-precision quality, with tool calling, thinking and a 256K context.

Prism ML's 9B vision model trained for ternary weights (-1, 0 or +1), about half the size of a standard 4-bit model at close to full-precision quality, with tool calling, thinking and a 256K context.

NVIDIA's compact vision-language model (74B total, 9B active parameters), distilled from Nemotron 3.5 Super and tuned to fit a 24 GB NVIDIA GPU for coding, reasoning and tool use.

IBM's flagship 30B foundation model — strong reasoning across business, coding, and multilingual tasks.

IBM's 8B dense foundation model for general instruction-following, RAG, and code tasks, with native tool use.

IBM's compact 3B instruction-tuned model with multilingual support, tool use, and structured JSON output.

Deep Reinforce's 35B MoE self-improving model (MIT) with 256K context, image input, and native tool use, strong across reasoning, coding, and agentic tasks.

Deep Reinforce's 9B self-improving model (MIT) with 256K context, image input, and native tool use, tuned for reasoning, coding, and agentic work.

Alibaba's 27B multimodal Qwen 3.8 model with a 256K context — substantial gains in coding, research, and long-horizon agentic tasks, with native image and video understanding.

Meta Superintelligence Labs' 30B multimodal model for always-on local agents — tuned for tool use, long tasks, and failure recovery.

NVIDIA's 30B mixture-of-experts model with 3B active parameters, built as the execution layer for always-on agents.

Poolside's 33B (3B active) MoE agentic-coding model — full-precision BF16 weights.

Poolside's 33B (3B active) MoE model for agentic coding and long-horizon local software work, with native reasoning and tool use.

Poolside's 33B (3B active) MoE agentic-coding model — Q8 quantization for higher fidelity.

Deep Reinforce's flagship 35B self-improving agentic-coding model (MIT) with 256K context — state-of-the-art open-source coding for its size.

Deep Reinforce's 9B self-improving agentic-coding model (MIT) with 256K context and native tool use.

Cohere's compact code-specialized model tuned for fast local code generation, completion, and refactoring.

Google's 12B multimodal Gemma 4 model with strong text and image understanding.

Liquid's efficient 8B Liquid Foundation Model tuned for long-context reasoning on-device.

Google's 26B mixture-of-experts model with long context and strong multimodal reasoning.

Google's flagship 31B dense model with 256K context and top-tier text and image understanding.

Google's efficient 2B multimodal model supporting text, image, and audio input.

Google's efficient 4B multimodal model — stronger reasoning than E2B with text, image, and audio.

IBM's flagship 30B dense foundation model — strong reasoning across business, coding, and multilingual tasks.

IBM's compact 3B instruction-tuned model with multilingual support, tool use, and structured JSON output.

IBM's 8B dense decoder-only foundation model for general instruction-following, RAG, and code tasks.

NVIDIA's multimodal Nemotron mixture-of-experts model unifying text and image understanding for enterprise Q&A, summarization, and document intelligence.

Alibaba's 27B multimodal Qwen 3.6 model with a 256K context and strong vision-language reasoning.

Alibaba's flagship 35B multimodal Qwen 3.6 model for top-tier vision-language reasoning.

Liquid's 24B foundation model with efficient long-context reasoning.

Alibaba's compact 2B multimodal Qwen model with vision input and a 256K context — fast on-device text+image reasoning.

Alibaba's efficient 4B multimodal Qwen model with vision input and strong multilingual reasoning.

Alibaba's 9B multimodal Qwen model balancing vision-language reasoning with on-device performance.

Google's medical multimodal model fine-tuned for healthcare and biomedical tasks.

Ai2's fully open 32B model — open weights, data, and training recipe — with long chain-of-thought reasoning.

Mistral's 14B dense edge model (Apache 2.0) tuned for fast on-device reasoning and tool use.

Mistral's compact 3B dense edge model (Apache 2.0) for fast on-device text generation and tool use.

Mistral's efficient 8B dense edge model (Apache 2.0) for on-device chat, reasoning, and tool use.

Ai2's fully open 32B model with open weights, training data, and recipe.

Ai2's fully open 7B model with open weights, training data, and recipe.

OpenAI's open-weight 20B reasoning model (Apache 2.0) with configurable reasoning effort and native tool calling — runs on 16 GB.

Large medical variant for complex clinical reasoning and biomedical analysis.

Mistral's 24B instruction-tuned model with vision input, robust function calling, and a 128K context.

DeepSeek's 8B reasoning model distilled from Qwen3 (R1-0528) with strong math, code, and logic.

Alibaba's compact 4B instruction model with strong reasoning and multilingual capabilities.

Alibaba's ultra-compact 0.6B Qwen3 model — runs in-process via HuggingFace or through Ollama, with tool use.

Microsoft's 14B reasoning model fine-tuned for chain-of-thought on math, science, and code.

Google's compact 1B Gemma 3 model — runs in-process via HuggingFace or through Ollama.

Microsoft's compact 3.8B model with a 128K context and native function calling.

DeepSeek's compact 1.5B reasoning model distilled from Qwen2.5 — chain-of-thought on-device.

DeepSeek's 14B reasoning model distilled from Qwen2.5 with exceptional math and coding benchmarks.

DeepSeek's 32B reasoning model distilled from Qwen2.5 for top-tier chain-of-thought reasoning.

Microsoft's 14B state-of-the-art open model with strong reasoning, math, and code.

Meta's 1B instruction-tuned model — the smallest Llama, runnable in-process via HuggingFace or through Ollama.

Meta's compact 3B instruction-tuned model — fast on-device text generation.

Meta's 8B instruction-tuned model with 128K context and strong general reasoning.

OpenMOSS's open 5B text-to-image model, running fully offline on your own machine. Qwen3.5 9B reads the prompt and the FLUX.2 VAE decodes the image, in 50 steps at CFG 3.

Black Forest Labs' open-weight FLUX 3 image model, running fully offline on your own machine. It generates from a prompt in 16 steps and edits images you attach.

Sperid Labs' 3B image model, running fully offline on your own machine. It generates directly in pixel space with no VAE, and needs no API key or network once the weights are downloaded.

The 8-step Turbo checkpoint of Alibaba Qwen's open 7B image model, running fully offline on your own machine. It generates from a prompt in about 40 percent of the steps the base model takes and needs no API key or network once the weights are downloaded. Released under the Qwen Research License, which covers research and evaluation use only.

Alibaba Qwen's open 7B image model, running fully offline on your own machine. It generates from a prompt and edits images you attach, renders long passages of text inside the image from a far smaller model than the 20B original, and needs no API key or network once the weights are downloaded. Released under the Qwen Research License, which covers research and evaluation use only.

Compact 4B Flux diffusion model for fast, high-quality image generation.

High-speed image generation model optimised for low-latency creative output.

Alibaba Qwen's 20B image model, running fully offline on your own machine. Known for rendering long passages of text — English and Chinese — inside the image, and needs no API key or network once the weights are downloaded.

A faster, slimmed-down Whisper Large V3 that keeps most of its accuracy and returns timed segments for subtitles.

On-device English text-to-speech that runs entirely in the app, with ten preset voices and no API key.

A small on-device Whisper model that transcribes speech in the app with no API key, trading accuracy for speed.

A mid-sized on-device Whisper model that transcribes speech in the app with no API key, balancing speed and accuracy.

MiniMax's H3 video model running offline, with synchronised audio. The heaviest local entry by a wide margin: its text encoder alone is a 32B vision-language model.

Lightricks' 22B video model, distilled for fewer sampling steps. Generates synchronised audio alongside the video, and drives a Gemma 3 language model as its text encoder.

Alibaba's largest open Wan model, a two-expert mixture that swaps between a high-noise and a low-noise transformer partway through denoising. Highest quality of the local Wan entries and by far the heaviest.

Alibaba's compact 5B video model, running fully offline on your own machine. Handles both text-to-video and image-to-video, and needs no API key or network once the weights are downloaded.

Alibaba's smallest Wan video model, and the lightest local option. Runs at 480p on modest hardware and shares its text encoder with the other Wan entries, so it costs little extra once one of them is installed.