Get occasional email updates on new AI tools. No spam.
"Open" means different things. Some of these ship their weights under a genuinely permissive license (Apache 2.0, MIT). Others use the word loosely — commercial-use caps, non-commercial-only weights, or "free" tools that aren't open source at all. We say which is which on every card.
The frontier keeps moving between "open" and "closed." OpenAI, once famously closed, shipped gpt-oss in 2025 — its first open weights in years. Meta, the open-source standard-bearer since Llama, now keeps its newest frontier model (Muse Spark) closed while still releasing smaller open ones like Muse Glimmer 30B. Read the license on every card below before you build on it — "open" doesn't always mean what you'd expect.
The last Llama release we could verify (April 2025) — natively multimodal, mixture-of-experts, very long context. Meta's newer frontier model, Muse Spark, is closed-weights; its newest open-weight model is Muse Glimmer 30B, below.
Europe's leading open-weight family. Medium 3.5 (spring 2026; sources disagree on the exact release date) is Mistral's newest flagship, under a modified MIT license — read the added terms. Mistral Large 3 and the small Ministral 3 models remain Apache 2.0.
One of the broadest open size ladders available, from laptop-friendly 27B up to frontier-scale MoE — the largest tier ships under a separate commercial license above $50M revenue.
DeepSeek's current main model (September 2026). It replaced V4-Pro — requests to V4-Pro are routed to V4.1-Flash from September 14. MIT license; takes text and image input. (We read the license from a Hugging Face discussion page, not the model card.)
Google's small-model line, now under a genuinely open license — the 31B dense variant reportedly ranks top-3 on LMArena's text leaderboard.
Small-footprint models tuned for reasoning-per-parameter efficiency — MIT licensed, runs comfortably on a single consumer GPU.
The rare model that's open end-to-end — training data, code, and intermediate checkpoints are public, not just the final weights, making results fully reproducible.
A hybrid attention/state-space architecture from the UAE's Technology Innovation Institute, claiming best-in-class reasoning for a 7B model.
OpenAI's first open weights in years — adjustable reasoning effort, quantized to run the 20B version on a consumer GPU with ~16GB of memory.
Built for long-horizon coding and agentic tasks with a 1M-token context — free to self-host, but large deployments need a commercial agreement and public attribution above set revenue/usage thresholds.
Same base as GLM-5.2 with post-training gains (August 2026). Worth knowing: 5.2 was MIT-licensed, but 5.3 moved to a custom license — check the terms before commercial use. Z.ai's docs page doesn't mention open weights; the Hugging Face repo does.
Meta's newest open-weight model — text and image in, text out. Meta's frontier Muse Spark stays closed, so this is where its open releases continue after Llama 4.
NVIDIA's Ultra-tier open-weight Nemotron 3 model (a 550B-parameter mixture-of-experts per its Hugging Face name), text only. A smaller Nemotron 3.5 Lightning (30B) followed on August 11. We haven't classified this license yet — read it before commercial use.
IBM's open Granite family, succeeding Granite 4.1. Text models; speech and vision are separate Granite models.
Cohere's flagship, now open-weight: one mixture-of-experts model that consolidates Command A Reasoning, Vision and Translate. 128K tokens in, 64K out.
MiniMax's flagship, succeeding M2.7. Takes image and video input (text output isn't stated). Weights were promised within about ten days of launch; we saw the Hugging Face repo but not its release date. Check the community license before commercial use.
An open vision-language model family aiming to match GPT-4o-class multimodal capability — image understanding, OCR, and visual reasoning. Current generation of the InternVL line (InternVL2 has been superseded).
A small vision-language model purpose-built to run efficiently on edge and low-resource devices — genuinely useful where a 78B model isn't an option. Weights update frequently under the same repo name, so pin a specific revision for production use.
Free for individuals and organizations under $1M annual revenue; larger orgs need a paid Enterprise license. The open-weight model that built the entire fine-tune/LoRA ecosystem around Automatic1111 and ComfyUI.
Built by ex-Stability researchers. Pick your variant carefully: only [klein] and the VAE are genuinely open-source; [dev] is open-weight but non-commercial by default, and [pro] is API-only.
The open-source standard for transcription and translation — permissively licensed, runs locally, and still nobody's officially topped it with a v4.
A tiny 82M-parameter model that punches far above its weight — 8 languages, ~54 voices, cheap enough to run on a laptop CPU.
Alibaba's open-weight answer to Stability's Stable Video Diffusion — genuinely open, no revenue caps, actively developed.
A drag-and-connect visual interface for building diffusion-model pipelines (image, video, audio generation) node by node — the standard way power users run and remix Stable Diffusion workflows. GPL-3.0 is copyleft, not permissive: modified/derivative code you distribute must also be released under GPL.
Retrieval-based Voice Conversion — trains a voice-conversion model from a short audio sample and converts speech or singing into that voice. Actively maintained by its community; Applio is a popular friendlier-UX fork built on the same technology, also MIT.
Clones a voice from a short reference clip and transfers it to text-to-speech output, including tone and emotion. V1 originally shipped non-commercial-only; both versions were relicensed to MIT in 2024, permitting commercial use.
Stability AI's 2023 base model — 3B parameters, two text encoders, usable alone or paired with a separate SDXL Refiner for higher-fidelity output. The inference code is MIT, but the weights ship under the RAIL++-M license, which carries use-based restrictions; it's since been overtaken by Stable Diffusion 3.5 for new work, though it remains the most fine-tuned and LoRA'd base model in the ecosystem.
A distilled version of SDXL that generates an image in a single network pass using Adversarial Diffusion Distillation — near-real-time speed at some cost to quality. Free to download and use non-commercially; commercial use needs a separate license from Stability AI.
Generates 1-second, 8-FPS animated GIFs designed to work alongside SDXL fine-tunes, LoRAs, and ControlNet. Worth knowing before you build on it: the repo hasn't been updated since January 2024 and appears abandoned, though it was never formally archived.
Adds a trained "motion module" on top of an existing Stable Diffusion checkpoint (SD1.5 or SDXL) to animate it, without retraining the base image model. This is the original research implementation; day-to-day development has largely moved to community integrations like ComfyUI-AnimateDiff-Evolved.
The easiest on-ramp for running most open-weight models on your own machine — one command to pull and run.
The low-level engine most local-AI tools (including Ollama) are quietly built on — runs GGUF-format models efficiently on modest hardware.
The production-grade choice for serving open-weight models at scale — used widely for self-hosted API endpoints.
A polished desktop app for running open-weight models fully offline, no terminal required — genuinely open source, not just free.
One of the most popular local-model GUIs — free for personal and commercial use, but the application itself is closed-source freeware, not open source, even though every model it runs is.
A local-first AI agent framework that manages messages, email, calendars, and files through a chat interface — your data stays on your machine, not a vendor's servers. Built and stewarded by an independent nonprofit; no paid or enterprise tier.
An open-source ChatGPT-style desktop app that runs models entirely offline — no account, no cloud call, works on Windows/macOS/Linux.
A browser-based UI for running, fine-tuning, and quantizing local models, with tool-calling and an OpenAI-compatible API — the power-user option once Ollama's simplicity isn't enough. AGPL-3.0 means a networked service built on a modified version must publish its source too, a stronger obligation than the copyleft on ComfyUI's GPL-3.0.
A simplified, offline SDXL interface built to remove manual prompt/parameter tuning. The original repo is now limited long-term support (bug fixes only, no new release since August 2024) and stays SDXL-only — no Flux or SD3.5 support planned. mashb1t/Fooocus ("1-Up Edition") is a community fork that's more actively maintained if you want ongoing updates.
The original, most-starred Stable Diffusion web interface, supporting SD1.x/SD2.x/SDXL/SD3 through a large extension ecosystem. Worth knowing: its latest tagged release dates to February 2025, and both users and extension developers have largely shifted toward ComfyUI and its forks (like Forge) since.
Chains together models, tools, and data sources to build LLM-powered applications and agents — one of the most widely adopted agent frameworks. The core library is fully MIT; LangSmith (tracing/eval/observability) is a separate paid, closed product built alongside it.
Connects, indexes, and retrieves your own documents/data so an LLM can answer questions grounded in them — the standard framework for building retrieval-augmented generation (RAG) systems. LlamaCloud (managed parsing/indexing) is a separate paid tier.
Builds modular LLM pipelines — RAG, semantic search, conversational systems — with explicit control over retrieval, routing, and generation steps. deepset also sells a separate paid, hosted "Haystack Enterprise" tier on top of the open-source core.
The standard Python library for loading, running, and training the actual model architectures behind most of what's on this page — text, vision, audio, and multimodal. Distinct from the Hugging Face Hub platform (listed separately below), which hosts paid tiers; this library itself is fully open source.
An open platform for autonomous coding agents that can modify code, run commands, browse the web, and call APIs to carry out real software-engineering tasks — renamed from OpenDevin. A separate paid OpenHands Cloud / Enterprise tier exists alongside the open-source core.
Rebuilt in 2026 as a Rust-based terminal coding agent optimized for low-cost/open-weight models — a real relaunch, not just a rename, of the original Python "let an LLM run code on your machine" project (that codebase is now an unofficial community fork, use with caution).
One of the earliest autonomous-agent experiments, now rebranded as a low-code agent-building platform with a marketplace. Important nuance: the original "classic" agent stays MIT, but the actively-developed new platform ships under Polyform Shield — a source-available license, not genuine open source.
Orchestrates role-playing "crews" of collaborating autonomous agents for complex, multi-step task automation. A separate paid Control Plane / AMP Suite (managed deployment, observability, governance) sits alongside the open-source MIT core.
A lower-level framework for building stateful, controllable, long-running agents and multi-step workflows as graphs — from the LangChain team, but usable standalone. LangSmith and a hosted deployment platform are separate paid products layered on top.
A polished, extensible self-hosted UI for chatting with local (Ollama) or cloud models, with RAG and access controls. Since v0.6.6 (April 2025) it's not plain open source: you can't remove/alter "Open WebUI" branding unless you're under 50 users/30 days or hold an enterprise license — a real, debated restriction, not OSI-approved.
An all-in-one app for building a private, ChatGPT-like assistant over your own documents — multi-user, pluggable LLM/vector-DB backends, built-in agents. Mintplex Labs also offers a separate paid hosted version alongside the free MIT core.
Blender Foundation's own Model Context Protocol connector — lets an MCP client like Claude query a Blender scene, rename/fix objects, explain material relationships, and generate documentation via natural language. It's built and maintained by Blender itself through their "Blender Lab" experimental program, distinct from earlier third-party Blender MCP add-ons. Blender's own docs warn it runs LLM-generated code with no sandboxing — use a VM, not your main machine.
Lets an AI assistant drive Krita over MCP — manage layers, draw, apply filters, export, and read the canvas back so the assistant can see what it painted. This is not a Krita Foundation project — it's an independent, single-maintainer tool with a small following (~35 GitHub stars at last check), so treat it as early-stage rather than production-ready.
Connects an AI assistant to Audacity for hands-free editing, cleanup, and mastering. Not affiliated with Audacity or Muse Group — it's an independent, single-maintainer project. Its own docs claim 144 tools across 9 pipelines; that count comes only from the maintainer's release notes, not an independent audit, so treat it as a self-reported figure.
A high-performance vector database built for large-scale approximate nearest-neighbor search — the retrieval backbone behind many RAG systems. Originally built by Zilliz, now a graduated Linux Foundation project; Zilliz Cloud is a separate paid managed tier.
Combines vector similarity search with structured filtering and hybrid search for RAG pipelines. Weaviate B.V. sells a separate paid Weaviate Cloud tier alongside the free open-source core.
An embedding/vector database built specifically for AI retrieval workflows — a common lightweight default for local RAG prototyping. Chroma Cloud is a separate paid managed offering on top of the open-source core.
A library (not a full hosted database) for efficient similarity search and clustering of dense vectors, CPU or GPU — a low-level building block often embedded inside other vector databases. No paid tier; Meta doesn't sell a hosted version.
A streamlined, config-file-driven way to fine-tune open LLMs without writing custom training code — one of the most widely used fine-tuning frameworks in the open community.
Hugging Face's official library for post-training language models — supervised fine-tuning and reinforcement-learning-based alignment methods like DPO and PPO.
Provides data/tensor/pipeline parallelism and memory optimization for training large models across many GPUs. Development cadence has slowed from its 2023-24 peak but the project remains maintained, not abandoned.
A deep-learning optimization library (ZeRO optimizer, mixture-of-experts support, inference kernels) used to train and serve very large models efficiently. The repo has moved from Microsoft's own GitHub org to an independent deepspeedai org — link the current one.
Deploys trained models (PyTorch, TensorFlow, ONNX, TensorRT, and more) in production across cloud and edge. Not the same project as OpenAI's unrelated "Triton" GPU-kernel language — a real naming collision worth knowing about if you go looking for it.
Builds optimized inference engines for running LLMs fast on NVIDIA hardware. The TensorRT-LLM code itself is Apache 2.0, but it wraps NVIDIA's closed-source TensorRT SDK under a separate proprietary license — and it's NVIDIA-GPU-only, not vendor-neutral.
Optimizes and deploys deep-learning inference across Intel hardware — the non-NVIDIA counterpart to TensorRT-LLM for teams running on CPUs or Intel accelerators rather than GPUs.
The de facto home for open-weight models and datasets. The Hub itself is a commercial product with paid tiers; its widely-used libraries are separately open-sourced under Apache 2.0.
The main gathering place for community fine-tunes and LoRAs built on Stable Diffusion and FLUX — the site itself isn't an open-source project, but almost everything hosted on it is freely downloadable.
Where the source code behind almost everything on this page actually lives. Not itself an “open source AI” project — just the shared workshop the ecosystem builds in.
License terms vary by model, and vendors don't always use "open source" precisely. Here's what each badge on this page stands for.
Permissive Apache 2.0 or MIT — free to use, modify, self-host, and build commercial products on, with no revenue caps or usage restrictions.
Restricted Weights are downloadable, but a custom license adds real conditions — revenue caps, MAU thresholds, or an acceptable-use policy layered on top.
Mixed The project ships multiple variants under different terms — some genuinely open, others not. Check which specific version you're pulling.
Free, not open source No cost to use, but the source isn't published under an open license — you can't inspect, fork, or self-host the actual application code.
Copyleft Genuinely open (e.g. GPL-3.0), but not permissive — if you distribute a modified version, you must release your changes under the same open license too. Different obligations than Apache 2.0/MIT, not a lesser license.
The term gets used loosely. In practice, "open source AI" today splits into six real categories — and the model you pick depends on which one you're actually getting.
1. Fully permissive (Apache 2.0, MIT) — no usage caps, no revenue limits, full commercial rights. Weights, and sometimes training data, are yours to use freely. On this page: DeepSeek-V4.1-Flash, Phi-4, OLMo 3.1, gpt-oss, Muse Glimmer 30B, Granite 4.2, Command A+, Mistral Large 3 / Ministral 3.
2. Open weights, restricted license — downloadable weights, but real conditions layered on top: an acceptable-use policy, a revenue cap, or a community license. On this page: Falcon-H1R (Apache-based + AUP), Kimi K3 (custom license, commercial terms above set thresholds), Llama 4 (Llama Community License), GLM-5.3 (custom license — 5.2 was MIT), MiniMax M3 (community license), Mistral Medium 3.5 (modified MIT).
3. Mixed licensing — some sizes or variants ship genuinely open, others under separate commercial terms. On this page: Qwen3.8 — Apache 2.0 up to its standard sizes, a custom commercial license kicks in above $50M revenue for the largest MoE tier.
4. Free, but not open source — no cost to use, but the source code isn't published under an open license, so you can't inspect, fork, or self-host the application itself. On this page: LM Studio — every model it runs is open, the app that runs them isn't.
5. Closed frontier, open small variants — a lab keeps its best model closed while still shipping smaller open releases. The clearest example right now: Meta's newest frontier model, Muse Spark (closed weights), sits alongside Muse Glimmer 30B (Apache 2.0, August 2026), its newest open-weight model. OpenAI moved the opposite direction with gpt-oss, its first open weights in years.
6. Copyleft (GPL-family) — genuinely open, but not permissive: distribute a modified version and you must release your changes under the same license too. On this page: ComfyUI (GPL-3.0) — different obligations than Apache 2.0/MIT, not a lesser license.
Open weights are only useful once something can actually run them. Three environments cover almost every case.
A) Locally, on a laptop or desktop — the right call for smaller, efficient models: Phi-4, Ministral 3's 3B–14B sizes, and gpt-oss-20b (built to run on a consumer GPU with ~16GB of memory) all fit comfortably. Run them through the tools in the "Run It Locally" tab above — Ollama, LM Studio, or GPT4All are the easiest on-ramps.
B) Cloud GPU providers — necessary for the frontier-scale MoE models: Kimi K3 (2.8T total), Nemotron 3 Ultra (550B total) and the large DeepSeek and GLM models are well beyond consumer hardware. Rented GPU platforms (RunPod, Lambda Labs, and the major clouds all support this) are the standard way to self-host at that scale.
C) Hybrid — local for iteration, cloud for scale — the common pattern for developers and agentic workflows: prototype against a small local model (fast, free, private), then point the same code at a cloud-hosted larger model once it needs to go to production. vLLM (in the "Run It Locally" tab) is built for exactly this kind of self-hosted serving at either scale.
Open-weight models aren't just chatbots — several of the models above are specifically built for agentic, tool-using workflows.
Coding agents — DeepSeek-V4.1-Flash and GLM-5.3 are both tuned for reasoning and coding.
Long-horizon research agents — Kimi K3's 1M-token context window is purpose-built for tasks that need to hold an entire codebase or document set in memory across a long agentic run.
Small, local agents — Phi-4 and Ministral 3's smaller sizes are efficient enough to run an always-on local agent without renting a GPU.
Fully reproducible agents — OLMo 3.1 publishes its training data and intermediate checkpoints, not just final weights, so an agent built on it is auditable end-to-end, not just at the output layer.
If you're picking your first open model to try, start here.
For a laptop: Phi-4 (MIT) — small, efficient, reasoning-focused. Ministral 3 (Apache 2.0) — a capable all-around open model at a manageable size. gpt-oss-20b (Apache 2.0) — runs on a consumer GPU.
For the cloud: DeepSeek-V4.1-Flash (MIT) — frontier-level reasoning at a fraction of typical training/serving cost. GLM-5.3 (custom license) — coding and agentic work; read its terms first. Kimi K3 — for long-horizon agent work needing its 1M-token context.
Six things actually decide which model is right for your project — in roughly this order.
License — Apache 2.0 or MIT is the safest default for commercial work; a restricted or mixed license (Falcon-H1R, Kimi K3, Qwen3.8's largest tier) means reading the actual terms before you ship.
Hardware — local (Phi-4, small Ministral 3) vs. cloud GPU (DeepSeek-V4.1-Flash, GLM-5.3, Kimi K3) is mostly a function of parameter count.
Use case — coding and agentic work favor DeepSeek-V4.1-Flash/GLM-5.3; broad general use favors Mistral or Llama 4; small-footprint reasoning favors Phi-4 or Falcon-H1R.
Context length — Kimi K3's 1M tokens stands apart; check the "Latest" quick-fact on each card above for the rest.
Efficiency — Phi-4 and Ministral 3's smaller sizes are built to get the most reasoning per parameter, not just the highest raw score.
Reproducibility — OLMo 3.1 is the one model here where training data and checkpoints, not just final weights, are public.
Local AI tooling is generally safe, but it's still software you're installing from many different maintainers — a few habits go a long way.
Always check the license before deploying commercially (see the badge guide above). Run unfamiliar tools in an isolated environment — a container or a VM — rather than directly on your main computer, especially anything that asks for elevated permissions. Prefer Apache 2.0 / MIT models for commercial work to avoid license surprises later. Keep models and inference tools updated, the same as any other software. And for frontier-scale models (Kimi K3, Nemotron 3 Ultra, the large DeepSeek and GLM models), a rented cloud GPU keeps the heavy compute — and any unfamiliar dependencies — off your personal machine entirely.
Expect more labs to adopt mixed licensing rather than a clean open/closed split — Qwen3.8's size-based split is likely a preview of the norm, not an exception. Small models keep getting disproportionately more capable (Phi-4, Gemma 4), which pushes serious local AI further into reach of ordinary laptops. Agentic, tool-using workflows are becoming the default way these models get used, not a side feature. And OLMo 3.1's fully-public training pipeline — data and checkpoints, not just weights — is a strong candidate for what "genuinely open" starts to mean by default, rather than the exception it is today.