// Open-weight models
Get open-weight models
Official repositories of open-weight models; free APIs and benchmarks are in the Models section.
53 tools51 open sourcechecked 2026-10-22Compare side by side →
How to choose · start with these three
53 / 53 items
Hugging Face The largest hub for open models, datasets and AI demo apps (Spaces). Closed / commercial Freemium WebAPI DeepSeek-V3 v1.0.0 ★ 105k Open source Qwen ★ 22k The official repo of Qwen (通义千问) chat & pretrained large language model proposed by Alibaba Cloud. Open source Weights MiniMax MiniMax's multimodal model platform for text, video (Hailuo), voice and music. Closed / commercial Freemium WebAPI LTX-2 ★ 9.8k Official Python inference and LoRA trainer for LTX-2. Open source Python cosmos ★ 12k NVIDIA Cosmos open platform of world models, datasets and tools. Open source platform rf-detr ★ 9.8k Real-time object detection and segmentation model architecture. Open source Python TabPFN ★ 8.2k Foundation model for tabular data. Open source Python ultralytics v8.4.175 ★ 63k Ultralytics YOLO27, YOLO26, YOLO11, YOLOv8 — object detection, instance segmentation, semantic segmentation, image classification, pose estimation, object tracking Open source SDKCLI Qwen3 ★ 28k Qwen3 is the large language model series developed by Qwen team, Alibaba Cloud. Open source Weights generative-models 0.1.0 ★ 27k Generative Models by Stability AI Open source Weights MiniCPM-V 20250527 ★ 27k A Pocket-Sized MLLM for Ultra-Efficient Image and Video Understanding on Your Phone Open source Weights DeepSeek-Coder ★ 24k DeepSeek Coder: Let the Code Write Itself Open source Weights gpt-oss v0.0.9 ★ 20k gpt-oss-120b and gpt-oss-20b are two open-weight language models by OpenAI Open source Weights Qwen3-VL ★ 20k Qwen3-VL is the multimodal large language model series developed by Qwen team, Alibaba Cloud. Open source Weights sam2 ★ 20k The repository provides code for running inference with the Meta Segment Anything Model 2 (SAM 2), links for downloading the trained model checkpoints, and example notebooks that show how to use the model. Open source Weights Wan2.2 2.2.11 ★ 18k Wan: Open and Advanced Large-Scale Video Generative Models Open source Weights Wan2.1 ★ 17k Wan: Open and Advanced Large-Scale Video Generative Models Open source Weights Qwen3-Coder ★ 17k Qwen3-Coder is the code version of Qwen3, the large language model series developed by Qwen team. Open source Weights dinov2 ★ 13k PyTorch code and models for the DINOv2 self-supervised learning method. Open source Weights HunyuanVideo HunyuanVideo ★ 13k HunyuanVideo: A Systematic Framework For Large Video Generation Model Open source Weights MiniCPM 5.0 ★ 12k MiniCPM5: SOTA on-device LLMs, small yet powerful. Open source Weights LTX-Video ltx-video-0.9.1 ★ 11k Official repository for LTX-Video Open source Weights Kimi-K2 ★ 11k Kimi K2 is the large language model series developed by Moonshot AI team Open source Weights ERNIE ernie-4.5 ★ 7.7k The official repository for ERNIE 4.5 and ERNIEKit – its industrial-grade development toolkit based on PaddlePaddle. Open source Weights llama-models v0.2.0 ★ 7.7k Utilities intended for use with Llama models. Open source Weights ml-fastvlm ★ 7.4k This repository contains the official implementation of "FastVLM: Efficient Vision Encoding for Vision Language Models" - CVPR 2025 Open source Weights InternLM v0.2.1dev20240102 ★ 7.3k Official release of InternLM series (InternLM, InternLM2, InternLM2.5, InternLM3). Open source Weights OLMo v0.6.0 ★ 6.7k Modeling, training, eval, and inference code for OLMo Open source Weights gemma v4.0.1 ★ 5.8k Gemma open-weight LLM library, from Google DeepMind Open source SDK GLM-4.5 ★ 4.4k GLM-4.5: Agentic, Reasoning, and Coding (ARC) Foundation Models Open source Weights Qwen3.8 ★ 4.3k The latest generation of Alibaba’s open Qwen models (27B posted on HF 2026-08-05, Flash-Next 2026-08-24): Qwen3.5-style hybrid Gated DeltaNet + gated attention with multi-token prediction, native image/video understanding, thinking on by default (switchable per request, depth via reasoning_effort), 262K native context extendable to 1M. The 27B dense model is Apache-2.0 and runs on Transformers, vLLM and SGLang; one of the most-liked models on HF (18k+) Open source PhiCookBook ★ 3.9k This is a Phi Family of SLMs book for getting started with Phi Models. Phi a family of open sourced AI models developed by Microsoft. Phi models are the most capable and cost-effective small language models (SLMs) available, outperforming models of the same size and next size up across a variety of language, reasoning, coding, and math benchmarks Open source Docs MiniMax-M2 ★ 2.6k MiniMax-M2, a model built for Max coding & agentic workflows. Open source Weights Step-Audio ★ 43 Open source Weights Aleph Alpha Kolibri 1 Aleph Alpha's 78B MoE reasoning model (3.46B active) released 2026-10-03, focused on German and English, with reasoning effort levels and tool calling, 256K native / 1M validated context; ~78 GB FP8 weights (1×H200/B200 or 2×H100), served via the aleph-alpha-inference vLLM plugin; EU GPAI Code of Practice signatory Open source AutoTrust JEV-35B AutoTrust's open 'System 1' decision model on Qwen3.5-35B-A3B (not affiliated with TypeSafe's closed Jev): up to 256 options per pass, served with vLLM + POST /v1/decide; ~72 GB bf16, NVFP4 variant 25.6 GB fits one 32–48 GB GPU Open source Cloudflare Clef Cloudflare's open 27B multimodal decision model: give it a state plus typed questions and it returns option probabilities in one forward pass (no free text); Jev/SystemOne API compatible, serve with SGLang; 9B Clef-Flash variant Open source DeepSeek V4.1 Flash 1017 2026-10-17 minor update of DeepSeek V4.1 Flash (open weights, MIT, image input). The official deepseek-flash API now serves it at unchanged prices: peak ¥2/¥8 per 1M input/output, cache hit ¥0.04, 50% off off-peak. #1 on OpenRouter trending this week. Model card: Terminal-Bench 4.0 42.3 (vs 31.2 for V4.1 Flash) Open source inclusionAI Ling 3.1 Flash inclusionAI Ling 3.1 Flash open weights (MIT); also listed on OpenRouter (~$0.075/$0.225 per 1M) Open source JetBrains Mellum 2.1 Thinking JetBrains code-focused MoE (12B-A2.5B) thinking variant; Apache-2.0 on the HF model card; published LiveCodeBench v6 / HumanEval+ numbers Open source Laya Convai Innovations non-autoregressive System 1 decision model (~0.4B): typed answers with calibrated probabilities in one pass; Jev-compatible /v1/systemone; English / multilingual / typed-decisions checkpoints Open source LFM2.5-24B-A2B Largest LFM2.5 model from Liquid AI: 24B total / ~2.45B active MoE for on-device agentic reasoning; 131K context; tool use for local agent harnesses Open source Liquid AI d1-3B Liquid AI's 3B multimodal decision model built on LFM2.5-VL-3B: give it a state (text/JSON/images) plus yes-no, choice or score questions and it returns calibrated probabilities in one pass with zero output tokens; ~8 ms per decision on an RTX 4090, ~30 ms on Apple M5 Pro. For routing, moderation, intent, agent guardrails — not a chat model Open source MiniCPM-V 4.7 OpenBMB MiniCPM-V latest (2026-10-12): first open MoE multimodal model tuned for smooth on-device use; 35B-A3B and 1B variants; single/multi-image and video understanding Open source NVIDIA Nemotron 3.5 Bolt 74B-A9B NVIDIA’s vision-language model for running everyday agentic tasks locally: Mamba-2/Transformer hybrid LatentMoE, 74B total / 9B active, up to 256K context, text/image/video in; native 2-bit (QAT) packing fits 24GB, deployed with vLLM on A100 / H100 / Blackwell per the card. Commercial use allowed (OpenMDW-1.1) Open source NVIDIA Nemotron 3.5 Super VL 120B-A12B NVIDIA's vision-language model: Mamba-2 + MoE + attention hybrid with a C-RADIO vision encoder, 120B total / 12B active, up to 1M context, text/image/video in; the NVFP4 build fits 1× DGX Spark or 1× H200; commercial use allowed (OpenMDW-1.1). On OpenRouter from $0.10/$0.40 per 1M, plus a free variant Open source Ornith-1.5 Ornith AI’s open models built on Qwen3.5 / Gemma4 with continued pre-/mid-/post-training, extending self-improvement to generating tasks, building scaffolds and RL on rollouts; the 9B dense model fits one GPU (with a mobile quantised variant), 35B-A3B is the MoE version; official GGUFs. MIT (check base-model terms too) Open source Perplexity Decider V1.2 27B / Flash V1.2 4B Perplexity's decision models (open weights, Apache-2.0): V1.2 27B (Qwen3.8-27B fine-tune) and Flash V1.2 4B (Qwen3.5-4B). Send a state plus yes-no / choice / score questions, get typed answers with probabilities; V1.2 reads all questions about one state in one sequence, so extra questions cost little. Official Decisions API (POST /v1/decisions): $0.02 per 1M input tokens ($0.01 Flash), output free; self-host or use OpenRouter Open source Reflection Beam 1.0 501B-A23B Reflection's open-weight reasoning model: sparse MoE, 501B total / 23B active, 1M context, mainly English; aimed at token-efficient coding and long-horizon agents (incl. MCP tools); BF16 / FP8 / NVFP4 weights. On OpenRouter since 2026-10-16 at ~$0.25–$1.3 input / $1.0–$4.3 output per 1M depending on provider Open source Ternary Bonsai 2 27B Prism ML’s ternary (−1/0/+1) build of Qwen3.8-27B: ~1.72 bits per language weight, GGUF ~5.95–7.21GB, the card claims 98.2% of FP16 quality and 262K context; runs on its llama.cpp fork (CUDA / Metal / CPU), with an MLX build for Apple silicon. Apache-2.0 Open source Underdog Saluki 27B 2-bit GGUF of Qwen3.8-27B (~7.89GB) tuned to keep tool calling for local agents; HF ConwayResearch/Underdog-Saluki-27B-1.0, Apache-2.0 Open source Xing4.0-29B-A4B (China Telecom TeleAI) China Telecom AI's Xing series (formerly TeleChat): MoE 29B total / 4B active, native 256K context (extensible to 512K), trained entirely on Ascend 910C; agent-oriented with adaptation for OpenCode, Claude Code, OpenClaw and Hermes; deploy with vLLM / SGLang / KTransformers. Free endpoint on OpenRouter Open source
About this data
- Default order is “recommended”: editorial picks first, then by popularity; general-purpose developer software and archived projects are listed last.
- Data checked 2026-10-22. Official links only. JSON API