A third category of AI model distribution has come to dominate the current landscape. Open-weight models release trained parameters for download while keeping the training data, code, and methodology proprietary. Six major labs now compete in this space, and these models account for the majority of non-API deployments in production. The term, and the specific trade-offs that it encodes, provide foundational context for the technical developments that this column tracks each week.

Parameters without the process

A neural network’s weights are the billions of numerical values that training produces. They encode the patterns that the model extracted from its data, from statistical relationships between words to factual associations and reasoning heuristics. A model the size of Llama 4 carries hundreds of billions of these values, and the file that contains them can run to several hundred gigabytes. When a lab publishes these parameters for download, it publishes the finished product. Anyone with sufficient hardware can run the model on private infrastructure, serve it through an internal API, or embed it in production software. The training data, the training code, and the methodology that shaped those parameters typically remain proprietary.

The term “open-weight” describes this middle position between fully closed models and fully open-source ones. Closed models, such as GPT-5 and Claude, are accessible only through an API, and the weights never leave the provider’s servers. Open-source projects, such as AI2’s OLMo, release the weights alongside the training data, training code, and evaluation methodology. Most major releases of the past year, including Llama, Qwen, Gemma, GLM-5, and gpt-oss from OpenAI, fall in the open-weight category, offering the parameters without the process that produced them.

Plug and play

Downloading the weights transfers control over deployment. An organization running the model inside its own network keeps proprietary data off third-party servers. Parameter-efficient techniques like LoRA allow fine-tuning on domain-specific data, adapting a general-purpose model to specialized tasks in hours on a single high-end GPU. Open-weight deployment also insulates infrastructure from provider decisions, eliminating API rate limits, unannounced pricing changes, and the risk that the lab deprecates the model on 60 days’ notice. Anthropic retired two Claude models from its API yesterday, and every application built on those model IDs had to migrate or break. Organizations running open-weight alternatives on their own infrastructure faced no equivalent dependency.

The training process, however, remains opaque. Without access to the training data, an organization cannot identify the model’s training sources, audit the corpus for bias, or assess copyright exposure. Without the training code and methodology, independent reproduction of the model’s capabilities stays out of reach. For regulated industries that require full auditability, including healthcare, finance, and defense, this opacity poses a genuine constraint. Most production deployments accept the trade-off because operational needs center on running and adapting the model.

A multi-polar landscape

Six major labs across three countries now ship frontier-competitive open-weight models, and permissive licensing has converged as the default. Google released Gemma 4 under Apache 2.0. Alibaba’s Qwen 3.6, OpenAI’s gpt-oss, and Mistral’s latest models all carry the same license. Zhipu AI published GLM-5, a 744-billion-parameter model trained entirely on Huawei Ascend chips without a single NVIDIA GPU, under MIT. That training run marked the first time that a frontier-class open-weight model achieved competitive benchmarks without any NVIDIA hardware in the pipeline. Only Meta’s Llama retains a custom community license. Epoch AI’s June data places the gap between the best open-weight models and the closed frontier at approximately four months on the organization’s Capabilities Index, a margin that held steady through the first half of 2026.

Meta, the company most identified with open-weight AI since Llama 2’s release in 2023, shipped its first closed proprietary model on April 8. Muse Spark launched across Facebook, Instagram, WhatsApp, and Meta’s smart glasses with no weights, no architecture paper, and no public methodology. The open-weight ecosystem absorbed the departure without destabilizing. Five other major labs had already established competitive positions, and the multi-polar structure meant that no single company’s participation was load-bearing. A year ago, Meta’s exit might have hollowed the field. In mid-2026, the departure registered as one data point in a landscape that had already diversified beyond any single point of failure.

Naturally anti-fragile

Open-weight release has established a durable category because it aligns the incentives of producers and deployers. Labs benefit from broad adoption and the downstream ecosystem that open weights create, while organizations gain sovereignty over their deployment stack. The landscape’s multi-polar structure reinforces this durability because no single lab’s strategic shift can destabilize a field in which five competitors offer viable alternatives under permissive licenses. The four-month gap to the closed frontier, meanwhile, establishes that open-weight models have entered the range of production viability for most applications. Each of those model families carries distinct architectural and strategic implications, and this column will examine them individually in the weeks ahead.

Subscribe now

AI Central

What “Open-Weight AI” Means