What open-weights models are — and why they matter
When an AI company releases a model, it has a choice: let you use it only through their website or API (closed), or publish the actual model files so anyone can download and run them (open weights). Open-weights models can be run on your own hardware, modified, fine-tuned for specific tasks, and built into products without paying per-query fees or worrying about a company changing its terms overnight.
For most of AI's modern history, the most powerful models were closed. That has changed dramatically.
How we got here: a quick history
The story starts with the foundational research — GPT-1 (2018) and GPT-3 (2020) showed the world what large language models could do, but the weights stayed proprietary. The open-weights movement got its real momentum from the community: BLOOM (2022), a 176-billion-parameter model built collaboratively by over 1,000 researchers, proved that frontier-scale open models were possible.
Then Meta changed the game. Llama 2 (July 2023) gave developers a genuinely capable open model they could download and run. Mistral 7B (September 2023) followed, outperforming models twice its size. Mixtral 8x7B (December 2023) introduced a clever trick — Mixture-of-Experts (MoE), explained below — that let a model be powerful but cheap to run.
From there, the pace accelerated into a full sprint.
The key players and their contributions
Meta / Llama is the anchor of the open-weights ecosystem. Llama 3 (April 2024), Llama 3.1 (July 2024, up to 405 billion parameters with multilingual support and long context), Llama 3.2 (September 2024, adding vision and tiny edge models), and Llama 4 (April 2025) each raised the bar. The Llama family is the most widely deployed open-weights lineage in history.
Mistral AI has shipped a relentless stream of models, almost all under the permissive Apache 2.0 license: Mistral 7B → Mixtral 8x7B → Mixtral 8x22B → Mistral Large 2 (123B) → Mistral 3 family (up to 675B total parameters) → Magistral (their first reasoning model) → Devstral (coding agents) → Mistral Small 4 → Mistral Medium 3.5. Mistral also built Le Chat and the Vibe CLI, turning its models into products.
Alibaba / Qwen has been equally prolific, releasing Qwen2 (June 2024), Qwen2.5 (September 2024, described as potentially the largest open-source release in history at the time), Qwen2.5-Coder (claiming parity with GPT-4o on coding), Qwen2.5-VL (vision-language), and Qwen3 (April 2025), whose 235-billion-parameter flagship claims competitive performance with DeepSeek-R1, OpenAI o1, and Gemini 2.5 Pro. The Qwen3-Coder 480B (July 2025) targets agentic coding specifically.
DeepSeek became a global story with DeepSeek-R1 — a reasoning model claiming parity with OpenAI o1, released under the MIT License (the most permissive possible), including six smaller distilled variants. DeepSeek-V3 (671B MoE, 37B active) followed, running at 60 tokens per second at remarkably low API prices. DeepSeek V4 (the current flagship) comes in a 1.6-trillion-parameter Pro variant and a 284-billion-parameter Flash variant, both with 1-million-token context windows by default.
Google DeepMind entered the open-weights race with Gemma 4 (April 2026), described as their most capable open models to date, purpose-built for reasoning and agentic workflows.
OpenAI — the company that started the modern LLM era with closed models — released gpt-oss-120b and gpt-oss-20b under Apache 2.0 in August 2025, a significant strategic reversal. The models are optimized for efficient deployment on consumer hardware.
Moonshot AI released Kimi K3 (July 2026), described as the largest open model ever released at 2.8 trillion total parameters (50 billion active), claiming performance comparable to Claude Opus 4.8 at a fraction of the price.
The clever trick: Mixture-of-Experts
Many of the largest open models use a technique called Mixture-of-Experts (MoE). Think of it like a hospital with many specialist doctors: instead of every doctor seeing every patient, a routing system sends each patient to the right specialist. Only a fraction of the model's total "knowledge" is activated for any given input, making the model fast and cheap to run despite having enormous total capacity. Mixtral 8x7B has 46.7 billion total parameters but only uses 12.9 billion for each token it processes — inference speed equivalent to a much smaller model.
The infrastructure that makes it all work
Models are only useful if people can run them. The open-weights ecosystem has built a rich stack:
- Hugging Face hosts nearly all open-weights models and provides the tools to download and use them. In February 2026, Hugging Face acquired llama.cpp and GGML — the foundational libraries that enable efficient local inference — securing their long-term development.
- vLLM, SGLang, and llama.cpp handle high-performance serving.
- LoRA and QLoRA (parameter-efficient fine-tuning techniques) let developers customize large models on modest hardware.
The tensions: safety, sovereignty, and distillation
Open weights come with real tradeoffs. Once model files are public, safety guardrails can be stripped out through fine-tuning. OpenAI studied this risk — called "malicious fine-tuning" — before releasing gpt-oss, specifically testing whether bad actors could elicit dangerous capabilities in biology and cybersecurity.
Anthropic identified a different risk: DeepSeek, Moonshot AI, and MiniMax allegedly ran large-scale "distillation attacks" — generating over 16 million exchanges through fraudulent accounts to train their own models on Claude's outputs. Anthropic framed this as a national security concern, arguing that illicitly distilled models strip out safety safeguards.
On the other side of the ledger, AI commentator Andrew Ng argued that restrictions on closed models — including U.S. export controls that forced Anthropic to disable global access to one of its models — are accelerating global interest in open-source alternatives. When a company or government can cut off access overnight, open weights become an insurance policy.
Where things stand
The gap between open and closed frontier models has narrowed to the point where, for many practical tasks, open-weights models are competitive. The remaining frontier — the very best reasoning, agentic, and multimodal capabilities — still tends to appear first in closed systems, but the lag has shrunk from years to months. The open ecosystem now has the models, the infrastructure, and the momentum to keep closing it.
