Open-Weight Models: Llama, Mistral, DeepSeek, and Qwen Explained

Open-weight AI models let you download, run, and fine-tune powerful LLMs on your own hardware - but "open" means different things for different models. Here's what builders need to know.

TL;DR: Open-weight models release the trained parameters (weights) so you can download and run them yourself. The big four families - Meta Llama, Mistral, DeepSeek, and Alibaba Qwen - each take a different approach to licensing, architecture, and "how open is open." This guide covers what matters for builders: what you can run, where you can run it, and what the fine print actually says.

What "Open-Weight" Actually Means

When a lab releases an open-weight model, it publishes the trained parameters - the billions of numerical values that make the model work. You can download them, run them on your own hardware, fine-tune them on your own data, and build products on top of them.

What you usually do not get: the training data, the full training code, or the pipeline used to produce those weights. That distinction matters enormously, and it is why the term open-weight is more accurate than "open-source" for most frontier models available today.

Open-Weight vs. Open-Source: A Real Difference

The Open Source Initiative (OSI) published its official Open Source AI Definition (OSAID) in October 2024. Under that definition, a truly open-source AI model must release weights, training data with enough detail to recreate the system, and training code - all under an OSI-approved license.

Almost no frontier-class model meets that bar today. OSI-validated models are mostly smaller research artifacts like Pythia (EleutherAI) and OLMo (AI2). Llama, Mistral, and DeepSeek release weights and some code, but not training data at scale. They are open-weight, not open-source in the strict sense - even when labs call them "open source" in press releases.

For most builders, this distinction is academic day-to-day. But it matters for compliance: if your legal or procurement team asks whether you are using "open-source AI," the honest answer for Llama or DeepSeek is "open-weight under a custom license."

Meta Llama: The Default Starting Point

Meta's Llama family is the most widely deployed open-weight series in the world. In April 2025, Meta released Llama 4 - the first natively multimodal generation, built on a mixture-of-experts (MoE) architecture.

The two released models are:

Llama 4 Scout - 17B active parameters, 16 experts, 109B total. Fits on a single NVIDIA H100 with int4 quantization. Industry-leading 10M token context window.
Llama 4 Maverick - 17B active parameters, 128 experts, 400B total. Runs on a single H100 DGX host (FP8 weights). 1M token context window. Benchmark-competitive with GPT-4o and Gemini 2.0 Flash.

A third model, Llama 4 Behemoth, was previewed but was still in training at release time - nearly 2 trillion total parameters.

The Llama License Fine Print

Llama does not use Apache 2.0 or MIT. It ships under the Llama 4 Community License Agreement - a custom license from Meta. Key points:

Royalty-free commercial use is allowed for most businesses.
If your product had more than 700 million monthly active users at the time of Llama 4's release, you must request a separate license from Meta (at Meta's sole discretion).
You must display "Built with Llama" prominently on related websites or product interfaces.
EU restriction - multimodal models only: For any multimodal models included in Llama 4, the rights granted are not available to individuals domiciled in, or companies with a principal place of business in, the European Union. Text-only Llama 4 variants are not subject to this restriction. EU end users can still access Llama 4 multimodal capabilities through a third-party service hosted outside the EU.
You can use Llama outputs to train other models (allowed since Llama 3.1), with attribution.

Bottom line: fine for most startups and projects globally. If you are EU-based and want to self-host a multimodal Llama 4 variant, use an Apache-licensed alternative instead.

Mistral: Lean, Fast, Apache-First

Mistral AI is a French lab with a reputation for punching above its weight class. Their open-weight models are notable for two things: they tend to be smaller and faster than comparably capable alternatives, and many ship under Apache 2.0 - one of the most permissive licenses available.

Key open-weight releases as of mid-2026:

Mistral Small 3 (24B, Apache 2.0) - latency-optimized, good for local or edge deployment.
Ministral series (3B, 8B, 14B) - efficient models with text and vision support, Apache 2.0.
Magistral Small (24B, Apache 2.0) - Mistral's first reasoning model, available for self-deployment. Builds on Mistral Small 3.1 with extended reasoning via SFT and RL. Fits on a single RTX 4090 or a 32GB MacBook when quantized.
Devstral Small 2 (24B, Apache 2.0) - code-focused, built for agentic coding tasks, 256K context window.
Mistral Large 3 (675B MoE, 41B active, Apache 2.0) - one of the most capable open-weight models available. Both base and instruct versions ship under Apache 2.0.

Where Mistral Gets Complicated

Not all Mistral models are Apache 2.0. Some ship under a modified MIT license with a revenue carve-out: companies generating more than $20 million USD in monthly revenue must obtain a commercial license from Mistral or access the model through Mistral Studio. Voxtral TTS uses CC BY-NC 4.0 (non-commercial only).

The rule: always read the model card for the specific model you are using. The licensing is per-model, not per-lab.

DeepSeek: The Efficiency Shock

DeepSeek, a Chinese AI lab backed by the quantitative hedge fund High-Flyer, made global headlines in January 2025 when DeepSeek-R1 matched OpenAI o1 on reasoning benchmarks - and released the weights under an MIT license.

MIT is as permissive as it gets: use it, modify it, sell products with it, fine-tune it. The only requirement is attribution.

DeepSeek-R1: What It Is

DeepSeek-R1 is a reasoning model - trained with reinforcement learning to "think" through problems step by step before answering, similar to OpenAI's o-series. The full model is 671B total parameters, 37B active (MoE), with a 128K context window. It matches o1 on math, code, and reasoning benchmarks.

Running 671B parameters requires serious hardware. DeepSeek solved this by releasing distilled variants - smaller models (1.5B, 7B, 8B, 14B, 32B, 70B) trained on R1's outputs, based on Qwen2.5 and Llama 3 architectures. The 32B distill gives near-frontier reasoning performance at a size that fits on a single consumer GPU.

In May 2025, DeepSeek released DeepSeek-R1-0528, an updated version with improved reasoning depth, reduced hallucinations, and added JSON output and function calling support - all still under MIT.

In April 2026, DeepSeek released DeepSeek-V4 (V4-Pro: 1.6T total parameters, 49B active, 1M token context; V4-Flash: 284B total, 13B active), also under the MIT license, continuing the lab's aggressive open-weight strategy.

The Distillation License Note

DeepSeek's distilled models are built on Qwen2.5 and Llama 3 base models. DeepSeek released the distilled weights under MIT, but the underlying base models carry their own licenses (Apache 2.0 for Qwen-derived, the Llama Community License for Llama-derived). In practice, you need to honor both: MIT (DeepSeek's grant) plus the base model's terms. The Llama-derived distills (8B, 70B) inherit the Llama 3.x Community License. Always check the model card for the specific variant you ship.

Qwen: Alibaba's Quiet Powerhouse

Alibaba Cloud's Qwen team has built one of the broadest open-weight ecosystems of any lab. Qwen3, released April 2025, offers dense models from 0.6B to 32B and MoE models at 30B-A3B and 235B-A22B. All open-weight Qwen3 models ship under Apache 2.0, with context windows up to 128K tokens.

Qwen3's key differentiator is a built-in thinking toggle: models can switch between a fast non-thinking mode and a slower reasoning mode that shows its work - useful when you want to dial the trade-off between speed and accuracy at inference time rather than swapping models.

By April 2026, the Qwen3.6 series extended context support to around 1 million tokens and maintained the Apache 2.0 license, making it the go-to choice for ultra-long-context work in the open-weight space. The models also support 100+ languages, making Qwen a strong default when you need multilingual coverage or want to fine-tune on non-English data.

One caveat: Alibaba's proprietary flagship models are available only through Alibaba Cloud. The open-weight lineup and the commercial-API lineup are separate products. Know which one you are evaluating.

How to Actually Run These Models

The tooling to run open-weight models locally has matured significantly. Two tools dominate:

Ollama - one-command install, pulls models from a registry, handles quantization automatically. Best for getting started fast.
llama.cpp - more control over quantization formats (GGUF), CPU/GPU split, and memory usage. Better for production self-hosting or constrained hardware.

A rough hardware guide at Q4 quantization (about 0.6-0.7 GB per billion parameters):

7B model - 6+ GB VRAM, runs on most modern consumer GPUs or Apple Silicon M-series.
32B model - 20-24 GB VRAM, fits on an RTX 4090/5090 or an M4 Max with 48GB unified memory.
70B model - needs 48+ GB VRAM, or CPU offload with enough RAM.
671B (full DeepSeek-R1) - requires a multi-GPU server or cloud inference.

For cloud inference, providers like Groq, Together AI, and Fireworks AI host most of the major open-weight models and charge per token - often cheaper than OpenAI for equivalent throughput, with no per-seat minimums.

# Pull and run Llama 4 Scout with Ollama
ollama pull llama4:scout

# Pull a DeepSeek R1 distill (32B, fits on RTX 4090)
ollama pull deepseek-r1:32b

# Pull Mistral Small 3
ollama pull mistral-small3

Choosing the Right Model for Your Project

A quick decision frame:

Need the most permissive license with no revenue thresholds? DeepSeek-R1 (MIT) or Qwen3 / Mistral Small 3 / Mistral Large 3 (Apache 2.0).
Need multimodal (images + text)? Llama 4 Scout or Maverick - the only frontier open-weight models with native vision at this tier. Note EU multimodal restriction for self-hosters.
Need strong reasoning at a size that runs locally? DeepSeek-R1-Distill-Qwen-32B or Magistral Small.
EU-based and want to self-host multimodal? The Llama 4 Community License restricts EU individuals and companies from deploying the multimodal models directly. Use Apache-licensed alternatives like Qwen3 (vision variants) instead.
Multilingual or non-English use cases? Qwen3 covers 100+ languages and often outperforms other models on non-English benchmarks.
Fast prototyping at minimal size? Mistral's Ministral 3B or Qwen3 0.6B run on CPU-only machines.

Key Takeaways

Open-weight means you get the trained parameters - not the training data or the full pipeline. It is not the same as open-source under the OSI definition.
Every major lab uses a different license. MIT (DeepSeek), Apache 2.0 (Qwen, most Mistral, Mistral Large 3), and custom community licenses (Llama) each have different implications for commercial use, EU entities, and revenue thresholds.
The Llama 4 Community License restricts EU-based individuals and companies from self-deploying the multimodal Llama 4 models specifically. Text-only variants and end-user access through third-party services are unaffected.
DeepSeek's distilled models are released under MIT by DeepSeek, but the Qwen-based distills also carry the Apache 2.0 terms of the underlying base model. Llama-based distills carry the Llama 3.x Community License. Check the model card before shipping.
Mistral's $20M monthly revenue threshold applies to certain models - not the whole catalog. Mistral Large 3 (675B) is Apache 2.0 with no revenue cap.
Qwen3 context windows top out at 128K. The 1M-token range arrived with Qwen3.6 in April 2026.
Ollama and llama.cpp are the two dominant local-inference tools. Cloud providers like Groq, Together AI, and Fireworks AI offer per-token pricing on most major open-weight models.

Try this next: Now that you know which model families exist and how the licenses work, see Inference Providers: Groq, Together AI, and Fireworks for a side-by-side comparison of hosted open-weight APIs - pricing, rate limits, and which models each provider carries.

Guide landscapecore9 min read