Open-Weight AI Models Compared

Thirteen labs now release AI models you can download and run on your own hardware. Meta, Alibaba, DeepSeek, Mistral, Google and Microsoft were joined through 2025 and 2026 by Z.ai, MiniMax, Moonshot, Cohere, OpenAI, IBM and Nous Research. Between them, they cover everything from a 3B parameter model that runs on a phone to Moonshot’s 2.8T parameter Kimi K3, which needs a data centre.

They now carry most production traffic on at least one large router. Vercel reports that open-weight models ran 56% of the tokens on its AI Gateway in August 2026, up from fewer than one in ten in December 2025 (Vercel, 17 September 2026).

But “open” carries a much narrower meaning than most people assume. And the modelModelA trained neural network that takes inputs (text, images, audio) and produces outputs (more text, classifications, generated content). In DeAI the model is the thing that actually does the work.Like a very experienced apprentice who has spent years watching thousands of masters make furniture. They can't explain how they know when a joint is right, but they can make a chair that looks and functions like a Chippendale. The training is invisible. The output is what matters.Model in the glossary → you choose determines what you can actually do with it, legally and practically. This page tracks the models that matter for sovereign inferenceInferenceRunning a trained AI model to produce an answer. Inference is what happens when you type a prompt into ChatGPT and get a response. The model takes your input, computes a best guess, and returns it.Like asking an expert for their opinion. The training was the decades they spent becoming an expert. The inference is the 30 seconds it takes them to answer your specific question.Inference in the glossary →: who made them, what licence they carry, how much hardware they need, and whether you can run them without asking anyone’s permission.

For the practical case for (and against) running these locally, see Why Self-Host Your AI. For hardware setup, see Mac Studio DeAI Setup.

Open-weight vs open-source: the distinction that matters

The AI industry uses “open source” loosely. Most models marketed as open source aren’t. The distinction is simple and it matters.

Truly open source means the model weightsParametersThe internal numbers (weights and biases) inside a neural network that get adjusted during training. A 70-billion-parameter model has 70 billion adjustable internal numbers encoding everything it has learned.Like the synapses in a human brain. Each parameter is a tiny dial that gets nudged a little during training. With enough dials, the network can represent surprisingly complex patterns. The total parameter count is roughly how much "brain" the model has.Parameters in the glossary → ship under a licence approved by the Open Source Initiative (OSI): Apache 2.0 or MIT. You can use the model commercially, modify it, redistribute it, fine-tune it, and deploy it without restriction. On this page that covers:

• Qwen (Alibaba): Qwen 3.5, 3.6 and 3.8
• DeepSeek V4-Pro and V4.1-Flash
• OpenAI gpt-oss-120b and gpt-oss-20b
• Mistral Large 3, Small 4 and Ministral 3
• Google Gemma 4, Apache 2.0 since 2026
• Z.ai GLM-5.3-Flash
• Cohere Command A+, which moves Cohere’s flagship off a non-commercial licence
• Microsoft Phi-4, IBM Granite 4.2 and Nous Research Hermes 4.3

Permissive, not OSI-approved covers NVIDIA’s Nemotron 3.5 Lightning. Its OpenMDW-1.1 licence lets you use, modify and redistribute the model without restriction, much like MIT. The difference is a termination clause: sue anyone claiming the model infringes a patent or copyright and your rights end. In practice it sits much closer to Apache 2.0 than to the restricted licences below.

Open weight, restricted licence means you can download and run the weights but the licence limits what you can do. Meta’s Llama models restrict use by applications with more than 700 million monthly active users. MiniMax’s community licence is non-commercial by default and gates larger commercial use behind written authorisation. Moonshot’s Kimi K3 licence is tighter than K2.6’s: a model-hosting business above US$20M in yearly revenue needs a separate agreement. These restrictions may not affect individual users today, but they constrain what you can build on top of them.

The trend runs both ways. Cohere and Google moved to Apache 2.0, while Moonshot tightened Kimi’s terms and Z.ai put its full GLM-5.3 under its own licence, keeping MIT for GLM-5.3-Flash. Check the licence on every release; the family name tells you nothing.

For sovereignty purposes, the difference matters. If you’re building infrastructure on top of a model, an Apache 2.0 or MIT licence means nobody can pull the rug. A restricted licence means someone can change the terms.

The comparison

Which model for which job

Hardware constraints and use case determine the right choice. Here’s how the current line-up maps onto common scenarios.

Laptop or Mini PC (8-16GB RAM). Ministral 3 3B, Qwen 3.5 9B, Gemma 4 12B, Phi-4, or gpt-oss-20b. All fit at Q4 quantisationQuantisationCompressing an AI model by storing each parameter with fewer bits of precision. Quantisation cuts model size and inference cost by 2-4x with small quality losses, making big models practical to run on consumer hardware.Like printing a high-resolution photo at lower DPI. The image is mostly the same, the details are slightly less crisp, and the file size drops dramatically. For most uses you can't tell the difference. For some uses the quality loss matters.Quantisation in the glossary →. Qwen 3.5 9B is the strongest all-rounder at this size, with Apache 2.0 licensing and native vision. gpt-oss-20b is a Mixture of ExpertsMoEMixture of Experts. A neural network architecture where many specialised "expert" sub-models exist alongside a router that picks which experts to use for each input. Only a fraction of the model's parameters are active for any single query.Like a hospital with many specialist doctors instead of one giant generalist. When a patient arrives, a triage nurse routes them to the right specialist. The hospital has 100 doctors total but each patient sees only the 2-3 they actually need.MoE in the glossary → model with only 3.6B parameters active per token, so it’s quick on modest hardware, at 14GB in Ollama. Phi-4 punches above its weight on reasoning but has a short 16K context window and is the oldest model here.

Mac Studio or workstation (32-64GB RAM). Qwen 3.8 27B is the sweet spot: about 18GB at Q4 and Apache 2.0. Gemma 4 31B (19GB) and IBM’s Granite 4.2 30B (18GB) are strong Apache 2.0 alternatives. Qwen 3.6 35B-A3B is the fast option, with only 3B parameters active per token, and NVIDIA’s Nemotron 3.5 Lightning (25GB in Ollama) matches it at 3B active with a 1M context window. Nous Research’s Hermes 4.3 36B (about 22GB) is the one here trained on a decentralised network, Nous’s Psyche. gpt-oss-120b fits on a 96GB+ machine at 65GB. Llama 3.3 70B fits at Q4 (42GB) but runs slowly enough to feel sluggish for interactive use.

Multi-GPU server or cloud. DeepSeek V4-Pro and V4.1-Flash are MIT-licensed with 1M context, but even V4.1-Flash is a 552B model. GLM-5.3-Flash (MIT, 320B total, 18B active) is Z.ai’s server-class option. Qwen 3.5 397B (Apache 2.0) and Mistral Large 3 (Apache 2.0) are the largest models here you can run with no licence restrictions at all, with Mistral as the European sovereignty option. Command A+ (Apache 2.0, 218B) gives Cohere’s enterprise line an open licence.

Coding and development. The server-class models (DeepSeek V4, GLM-5.3-Flash, Qwen 3.5 397B) are the open options for heavy coding work. For something that runs locally, Qwen 3.8 27B and Qwen 3.6 35B-A3B handle code generation well. For editor-integrated completion via Continue.dev, a 24-36B model gives responsive autocomplete.

RAG and document analysis. Context window matters here. Qwen 3.5 to 3.8, Gemma 4 and Mistral Small 4 support 256K tokens natively, and Hermes 4.3 512K. DeepSeek V4, GLM-5.3-Flash, MiniMax M3 and Kimi K3 reach 1M. For embeddingEmbeddingA numerical representation of a word, sentence, or image as a list of numbers (a vector) that captures its meaning. Similar things have similar embeddings, which makes them useful for search, clustering, and recommendation.Like a map where every word has GPS coordinates. Words with similar meanings end up close together on the map. "Cat" and "kitten" are nearby. "Cat" and "thunderstorm" are far apart. The map is the embedding space.Embedding in the glossary → generation, run nomic-embed-text or BGE-M3 alongside your generation model: they’re under 1GB and don’t compete for memory.

Privacy-first use cases. Any model running via Ollama or LM Studio on your own hardware provides Level 4 privacy (the highest on our privacy spectrum). No prompts leave the machine. For situations where data can’t touch an external server, local inference with an open-source model is the only option that provides both legal clarity and architectural guarantees.

The sovereignty lens

Three questions determine whether a model is sovereign-compatible:

1. Can you run it without permission? Apache 2.0 and MIT: yes. Llama Community, MiniMax Community and the Kimi K3 licence: conditionally.

2. Can you build on it without risk? If the licence allows commercial use and modification without restrictions, you can build products, fine-tune for clients, and redistribute. If not, your project depends on someone else’s licence terms staying favourable.

3. Can you run it on hardware you control? Models under 70GB at Q4 run on consumer hardware. Above that, you need data centre infrastructure, which may mean renting from a cloud provider and reintroducing a dependency.

The models that score highest on all three: Qwen 3.8 27B (Apache 2.0, 18GB, runs on a Mac), gpt-oss-20b (Apache 2.0, 14GB, runs on a laptop with 16GB), Gemma 4 12B (Apache 2.0, 8GB) and Qwen 3.5 9B (Apache 2.0, 6GB, runs on almost anything). DeepSeek’s and Z.ai’s MIT-licensed flagships are excellent but too large for consumer hardware.

How we keep this current

Open-weight models release constantly. This page uses a hybrid approach:

Automated data refresh. Download counts and community metrics update weekly via the HuggingFace APIAPIApplication Programming Interface. A structured way for one piece of software to talk to another. In DeAI, APIs let applications request inference from a model without running the model themselves.Like a waiter in a restaurant. You don't walk into the kitchen and cook your own meal. You tell the waiter what you want, they tell the kitchen, the kitchen cooks it, and the waiter brings it back. The API is the waiter.API in the glossary → (npm run refresh:models, run in CI every Monday). The “last refreshed” date at the top of the comparison shows when data was last pulled.

Manual editorial updates. New model families, licence changes, hardware requirements, and sovereignty assessments are updated during our monthly review process. We add models when they’re significant enough to change the comparison, not every time someone releases a 7B variant. Hardware figures are Ollama’s Q4 download sizes where Ollama carries the model; the largest models, which have no local Ollama build, use an estimate from their parameter count.

What we track. 24 models from fourteen labs, chosen for relevance to sovereign and local inference. We exclude models that are research-only or too niche to matter for our audience.

For how these models fit into a broader sovereignty strategy, see The Sovereignty Stack.