---
title: "Open vs closed AI models: the AI Model Release Report 2026"
description: "Open vs closed AI models in 2026: open-weight models fell to 45.9% of new releases from 60.1% a year earlier, across 1,250 releases from 73 AI labs."
summary: "Open-weight models made up 45.9% of new AI model releases from January 1 to November 16, 2026, down from 60.1% in the same period of 2025. This report breaks the open vs closed split down by year, model type, country, and lab, and shows that the drop holds however the data is cut."
type: "comparison"
canonical_url: "https://theaiengineering.com/articles/open-vs-closed-ai-models"
markdown_url: "https://theaiengineering.com/articles/open-vs-closed-ai-models.md"
topic: "LLM models and benchmarks"
author: "The AI Engineering Editors"
publisher: "The AI Engineering"
published: "2026-11-19"
updated: "2026-11-19"
keywords: ["open vs closed ai models","open weight vs closed weight models","open source vs closed source ai models","open weight models","open weight vs proprietary models","ai model releases 2026"]
citation: "The AI Engineering Editors. \"Open vs closed AI models: the AI Model Release Report 2026.\" The AI Engineering, updated Nov 19, 2026. https://theaiengineering.com/articles/open-vs-closed-ai-models"
---

# Open vs closed AI models: the AI Model Release Report 2026

## TL;DR

- Open-weight models made up 45.9% of new AI model releases from January 1 to November 16, 2026, down from 60.1% in the same period of 2025, across the 73 labs in The AI Engineering release tracker.
- Proprietary releases rose from 149 to 237 between the two periods, while open-weight releases fell from 224 to 201.
- The drop holds among the 45 labs active in both years (59.9% to 45.7%), for text models alone (70.4% to 58.7%), and with any single lab left out.
- US labs drove most of the shift: their open-weight share fell from 50.9% to 32.3%, while the share for Chinese labs moved from 66.4% to 63.5%.
- Labs headquartered in India released 37 models in the 2026 period, up from 21, and 18 of the 37 were open-weight.

Open-weight models made up 45.9% of new AI model releases between January 1 and November 16, 2026, down from 60.1% in the same period of 2025. That is the main finding of the AI Model Release Report 2026, which compares open vs closed AI models across 1,250 releases from 73 labs logged in The AI Engineering's [model release tracker](https://theaiengineering.com/releases) since January 2023. Every count below comes from that tracker, where each entry links to the lab's own announcement, and the full dataset is free to download.

## Open vs closed AI models in 2026: the headline numbers

From January 1 to November 16, 2026, the tracker logged 438 releases: 201 with open weights and 237 proprietary. In the same window of 2025 it logged 373: 224 open-weight and 149 proprietary. Proprietary releases grew by 88 while open-weight releases fell by 23, so the open-weight share fell 14.2 percentage points.

An [open-weight model](https://theaiengineering.com/glossary/open-weights-model) is one whose trained parameters the lab publishes for download. A proprietary model is offered only through the lab's API or apps. Each year below uses the same window, January 1 to November 16, so the years compare fairly.

| Year | Releases | Open-weight | Proprietary | Open share | Full-year share |
|---|---|---|---|---|---|
| 2023 | 85 | 46 | 39 | 54.1% | 53.4% |
| 2024 | 205 | 135 | 70 | 65.9% | 64.0% |
| 2025 | 373 | 224 | 149 | 60.1% | 60.6% |
| 2026 | 438 | 201 | 237 | 45.9% | In progress |

The tracker's first entry is dated January 30, 2023, so the 2023 window is slightly shorter. Open-weight releases outnumbered proprietary ones in every earlier year, which makes 2026 the first year in the data where proprietary releases lead. The 201 open-weight releases of 2026 still exceed the 158 logged in all of 2024.

## Where the growth in proprietary releases came from

US labs added 79 of the 88 extra proprietary releases between the 2025 and 2026 windows. Five of them account for most of it: Google, OpenAI, Microsoft, xAI, and Anthropic released 108 proprietary models in the 2026 window, up from 53. Indian labs added 17 proprietary releases, while Chinese labs and labs elsewhere released slightly fewer.

| Lab | Proprietary, 2025 window | Proprietary, 2026 window |
|---|---|---|
| Google | 20 | 34 |
| OpenAI | 21 | 28 |
| Microsoft | 3 | 19 |
| xAI | 4 | 15 |
| Anthropic | 5 | 12 |

Some of this growth is how labs ship. A preview and its final release count as two entries when the lab announces them separately, and Google announced 41 models in the 2026 window. Removing preview, experimental, and beta entries changes the 2026 share by less than half a point (45.5%), so the shift is not an artifact of preview releases.

## Open-weight share by model type

The open-weight share fell for text models and fell further for image, video, speech, and music models. Text models, which include [large language models](https://theaiengineering.com/glossary/large-language-model), reasoning, coding, [small](https://theaiengineering.com/glossary/small-language-model), and [multimodal models](https://theaiengineering.com/glossary/multimodal-model), went from 70.4% open-weight to 58.7%. Media models went from 42.7% to 25.0%.

| Model type | 2025, open-weight of all | 2026, open-weight of all |
|---|---|---|
| Text models | 159 of 226 (70.4%) | 145 of 247 (58.7%) |
| Image, video, speech, and music | 56 of 131 (42.7%) | 44 of 176 (25.0%) |
| Embedding and reranker | 9 of 16 (56.3%) | 12 of 15 (80.0%) |

Speech drove the media numbers. Speech releases nearly doubled, from 44 to 82, and 60 of the 82 in 2026 were proprietary. Image models went the same way: 16 of 36 were open-weight in 2025 and seven of 45 in 2026. [Embedding](https://theaiengineering.com/glossary/embeddings) models are the exception, though the counts are small.

## US, Chinese, and Indian labs

The drop is concentrated in US labs. Their open-weight share fell from 50.9% to 32.3% as their proprietary releases doubled, while Chinese labs held close to two-thirds open-weight, moving from 66.4% to 63.5%. Indian labs went from 19 open-weight releases out of 21 to 18 out of 37.

| Lab headquarters | 2025, open-weight of all | 2026, open-weight of all |
|---|---|---|
| United States | 81 of 159 (50.9%) | 75 of 232 (32.3%) |
| China | 99 of 149 (66.4%) | 80 of 126 (63.5%) |
| India | 19 of 21 (90.5%) | 18 of 37 (48.6%) |
| Everywhere else | 25 of 44 (56.8%) | 28 of 43 (65.1%) |

Shares within a country are steadier than head-to-head counts. Chinese labs released 80 open-weight models in the 2026 window and US labs 75, a gap small enough that adding labs the tracker does not yet cover, in either country, could close it.

## Labs that released the most open-weight models

Liquid AI released the most open-weight models in the 2026 window, 22 of its 24 releases. Alibaba's Qwen team and Tencent followed with 14 and 13, both fewer than in 2025, when they released 23 and 18. Most labs still publish some weights: 44 of the 61 labs active in 2026 released at least one open-weight model.

| Lab | Open-weight, 2026 | All, 2026 | Open-weight, 2025 | Based in |
|---|---|---|---|---|
| [Liquid AI](https://theaiengineering.com/releases/liquid-ai) | 22 | 24 | 8 | United States |
| [Alibaba Qwen](https://theaiengineering.com/releases/alibaba-qwen) | 14 | 25 | 23 | China |
| [Tencent](https://theaiengineering.com/releases/tencent) | 13 | 15 | 18 | China |
| [NVIDIA](https://theaiengineering.com/releases/nvidia) | 10 | 10 | 13 | United States |
| [IBM](https://theaiengineering.com/releases/ibm) | 9 | 9 | 12 | United States |
| [Meituan](https://theaiengineering.com/releases/meituan) | 9 | 10 | 3 | China |
| [Mistral AI](https://theaiengineering.com/releases/mistral) | 9 | 13 | 8 | France |
| [Z.ai](https://theaiengineering.com/releases/z-ai) | 9 | 11 | 7 | China |
| [Cohere](https://theaiengineering.com/releases/cohere) | 7 | 10 | 6 | Canada |
| [DeepSeek](https://theaiengineering.com/releases/deepseek) | 7 | 7 | 10 | China |

Google released 13 open-weight models in the 2025 window and seven in 2026, while its proprietary releases rose from 20 to 34. The number of labs with at least one open-weight release rose from 30 of 47 active labs in 2024 to 38 of 50 in 2025 and 44 of 61 in 2026.

## Indian labs in the tracker

The tracker includes 17 labs headquartered in India. In the 2026 window, 12 of them released 37 models, up from 21 releases by seven labs in 2025. [Sarvam AI](https://theaiengineering.com/releases/sarvam-ai) released 13, including its open-weight 30B and 105B models in March, and speech was the largest category with 13 releases.

| Lab | Releases, 2026 window | Open-weight |
|---|---|---|
| Sarvam AI | 13 | 3 |
| Gnani AI | 4 | 1 |
| Mannas | 4 | 3 |
| BharatGen | 3 | 3 |
| Krutrim | 3 | 3 |
| NPCI | 2 | 0 |
| Soket AI | 2 | 1 |
| ZenteiQ | 2 | 2 |
| ARTPARK (IISc), Avataar, Josh Talks, Tech Mahindra | 1 each | 2 of 4 |

Sarvam released [Sarvam 30B and Sarvam 105B](https://www.sarvam.ai/blogs/sarvam-30b-105b) with open weights on March 6, 2026, and 10 of its 13 releases in the window were proprietary, most of them speech and vision models. Indian labs released 19 open-weight models out of 21 in the 2025 window, so their share fell further than the global average.

## The drop holds in every cut of the data

The drop shows up in every cut of the data tested. Limited to the 45 labs that released models in both periods, the open-weight share fell from 59.9% to 45.7%. With any single lab left out, the 2026 share stays between 43.2% and 48.9%, and the 2025 share between 58.0% and 63.4%, so the two ranges never overlap.

| Cut of the data | 2025 window | 2026 window |
|---|---|---|
| All 73 labs | 60.1% | 45.9% |
| Only the 45 labs active in both periods | 59.9% | 45.7% |
| Text models only | 70.4% | 58.7% |
| Without the image, video, speech, and music labs | 63.0% | 48.6% |
| Without Google, Meta, Microsoft, Amazon, and Apple | 62.4% | 52.8% |
| Without the five busiest labs of 2026 | 64.7% | 52.0% |
| Without preview, experimental, and beta entries | 60.5% | 45.5% |

The five busiest labs of 2026 are Google, OpenAI, Alibaba Qwen, Liquid AI, and Microsoft. The exact share depends on which labs are counted. Without big tech, the 2026 share is 52.8% instead of 45.9%. Weighting every lab equally, the average lab's open-weight share fell from 59.2% to 51.3%, after rising from 52.2% in 2024.

## Choosing between open-weight and proprietary models

Fewer new models arrive with downloadable weights, so teams that need to run models on their own hardware have a smaller share of each month's releases to choose from. The absolute supply is still large: 201 open-weight releases in the 2026 window, more than in all of 2024, from 44 different labs.

Open weights matter when data must stay on your infrastructure, when you need to [fine-tune](https://theaiengineering.com/glossary/fine-tuning) or [quantize](https://theaiengineering.com/glossary/quantization) a model, or when you want to pin a version that an API provider might retire. The [self-hosted LLM guide](https://theaiengineering.com/articles/self-hosted-llm) covers the serving side, and [vLLM vs Ollama](https://theaiengineering.com/articles/vllm-vs-ollama) compares two common ways to run open-weight models. To compare open-weight and proprietary candidates on quality, start with how to read [LLM benchmarks](https://theaiengineering.com/articles/llm-benchmarks).

## Methodology

The data comes from The AI Engineering release tracker: 1,250 releases from 73 labs, dated January 30, 2023 to November 16, 2026, counted on November 17, 2026. Each record carries the lab's official announcement date and links to that announcement. Labs headquartered in India were added in November 2026, with entries back to 2023 checked against each lab's own pages.

| Item | How it was counted |
|---|---|
| Open-weight | The lab published the model's weights for download at or shortly after launch. This is close to how the NTIA's [July 2024 report on open model weights](https://www.ntia.gov/issues/artificial-intelligence/open-model-weights-report) describes widely available weights, and narrower than open source AI under the [Open Source AI Definition](https://opensource.org/ai/open-source-ai-definition), which also requires training code and data information. |
| Window | January 1 to November 16 of each year, the last date in the data, so 2026 compares with equal stretches of earlier years. |
| Release | A notable release a lab announced, not every fine-tune or checkpoint posted online. Variants announced separately, such as a preview and its final version, count separately. |
| Country | The lab's headquarters as stated on its own site. Fractal is in the tracker's India group but lists New York as its global headquarters, so it counts as a US lab. |
| Coverage | The 73 tracked labs. Shares and year-on-year changes are more reliable than head-to-head counts between countries or labs, which can move as coverage grows. |

This report is part of the [models topic hub](https://theaiengineering.com/topics/models), and the [2026 release archive](https://theaiengineering.com/releases/2026) lists every model counted here.

## Download the data

The dataset behind this report is free to download as [CSV](https://theaiengineering.com/releases.csv) or [JSON](https://theaiengineering.com/releases.json) under a [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/) license, which allows reuse with attribution. Each record lists the release date, lab, model, type, and weights status, with a link to the lab's own announcement. The tracker updates as labs announce new models, so later downloads will include releases after November 16, 2026.

## Frequently asked questions

### What is the difference between open and closed AI models?

An open-weight model publishes its trained parameters for download, so a team can run, fine-tune, or quantize it on its own hardware under the terms of its license. A closed, or proprietary, model is offered only through the lab's API or apps, and its weights stay private. Open weights is narrower than open source AI, which under the Open Source Initiative's definition also requires training code and information about the training data.

### What share of new AI models were open-weight in 2026?

Open-weight models made up 45.9% of the 438 releases The AI Engineering's tracker logged between January 1 and November 16, 2026. In the same period of 2025, the share was 60.1% of 373 releases, and in 2024 it was 65.9% of 205. The counts cover the 73 labs in the tracker, not every model posted online, and 2026 is the first year in the data with more proprietary releases than open-weight ones.

### Which AI labs released the most open-weight models in 2026?

Among the 73 labs in the tracker, Liquid AI released the most open-weight models from January 1 to November 16, 2026, with 22, followed by Alibaba's Qwen team with 14 and Tencent with 13. NVIDIA released 10, and IBM, Meituan, Mistral AI, and Z.ai released nine each. Google released the most models overall, 41, of which 34 were proprietary.

### Are Chinese AI labs more open than US labs?

By share, yes, in the tracker's data. From January 1 to November 16, 2026, Chinese labs released 80 open-weight models out of 126, or 63.5%, while US labs released 75 out of 232, or 32.3%. A year earlier the shares were 66.4% and 50.9%. The raw counts of open-weight releases are close, so adding labs the tracker does not yet cover could change which country leads on count.

### How many AI models did Indian labs release in 2026?

Labs headquartered in India released 37 models between January 1 and November 16, 2026, up from 21 in the same period of 2025, according to the tracker. Sarvam AI released 13 of them, and speech models were the largest group, with 13 releases. Of the 37, 18 were open-weight, compared with 19 of the 21 a year earlier.

### Can I reuse the AI model release data?

Yes. The dataset behind this report is free to download as CSV or JSON under a CC BY 4.0 license, which allows reuse, including commercial reuse, with attribution. Each record lists the release date, lab, model, model type, and whether the weights were published, and links to the lab's own announcement, so every count in this report can be checked.

## Terms in this guide

- [Open-weights model](https://theaiengineering.com/glossary/open-weights-model): An open-weights model is an AI model whose trained parameters are published for download, so anyone can run, fine-tune, or quantize it on their own hardware under the terms of its license.
- [Large language model (LLM)](https://theaiengineering.com/glossary/large-language-model): A large language model (LLM) is a neural network, usually a transformer with billions of parameters, trained on web-scale text to predict the next token and then adapted to follow instructions.
- [Small language model](https://theaiengineering.com/glossary/small-language-model): A small language model (SLM) is a language model compact enough to run on a consumer device such as a laptop or phone, which in 2025 generally meant fewer than about 10 billion parameters.
- [Multimodal model](https://theaiengineering.com/glossary/multimodal-model): A multimodal model is an AI model that takes in or produces more than one type of data, such as text, images, audio, or video, and combines them within a single model.
- [Reasoning model](https://theaiengineering.com/glossary/reasoning-model): A reasoning model is a large language model trained, usually with reinforcement learning, to generate intermediate reasoning tokens before its final answer, spending more inference compute to solve multi-step problems.
- [Embeddings](https://theaiengineering.com/glossary/embeddings): Embeddings are fixed-length vectors of numbers that represent data such as text or images so that items with similar meaning sit close together, which lets software compare meaning with arithmetic.
- [Fine-tuning](https://theaiengineering.com/glossary/fine-tuning): Fine-tuning is the process of further training a pretrained model on a smaller, task-specific dataset so its weights change to improve performance on a particular task, domain, style, or output format.
- [Quantization](https://theaiengineering.com/glossary/quantization): Quantization is a model compression technique that represents a neural network's weights, and sometimes its activations or KV cache, with fewer bits, such as 8-bit or 4-bit integers instead of 16-bit floats.
- [Foundation model](https://theaiengineering.com/glossary/foundation-model): A foundation model is a large AI model trained on broad data, usually with self-supervised learning at scale, that can be adapted through prompting or fine-tuning to a wide range of downstream tasks.

## Related guides

- [LLM benchmarks explained: what they measure and how to read them](https://theaiengineering.com/articles/llm-benchmarks): LLM benchmarks explained: what MMLU-Pro, GPQA, SWE-bench, LiveCodeBench, ARC-AGI, and LMArena measure, why scores saturate, and how to read a leaderboard.
- [How to self-host an LLM: hardware, serving engines, and a production setup](https://theaiengineering.com/articles/self-hosted-llm): Self-hosted LLM guide: size GPU memory for weights and KV cache, run models locally with Ollama or llama.cpp, and serve them with vLLM, SGLang, or Kubernetes.
- [Best LLM for coding: models compared on agentic coding benchmarks](https://theaiengineering.com/articles/best-llm-for-coding): The best LLM for coding depends on the job: 11 models compared on Terminal-Bench 4.0, DeepSWE, and SWE-bench Pro, plus context, open weights, and cost.
- [Reasoning models: how they work and when to use them](https://theaiengineering.com/articles/reasoning-models): Reasoning models explained: how test-time compute and reinforcement learning produce chain-of-thought reasoning, when to use them, and their cost and latency.

Part of the [LLM models and benchmarks](https://theaiengineering.com/topics/models) topic hub.

## Cite this guide

The AI Engineering Editors. "Open vs closed AI models: the AI Model Release Report 2026." The AI Engineering, updated Nov 19, 2026. https://theaiengineering.com/articles/open-vs-closed-ai-models

- Title: Open vs closed AI models: the AI Model Release Report 2026
- Author: The AI Engineering Editors
- Publisher: The AI Engineering
- Published: 2026-11-19
- Updated: 2026-11-19
- URL: https://theaiengineering.com/articles/open-vs-closed-ai-models
- Markdown: https://theaiengineering.com/articles/open-vs-closed-ai-models.md

To cite one section, link to its anchor:

- [Open vs closed AI models in 2026: the headline numbers](https://theaiengineering.com/articles/open-vs-closed-ai-models#open-vs-closed-ai-models-in-2026-the-headline-numbers)
- [Where the growth in proprietary releases came from](https://theaiengineering.com/articles/open-vs-closed-ai-models#where-the-growth-in-proprietary-releases-came-from)
- [Open-weight share by model type](https://theaiengineering.com/articles/open-vs-closed-ai-models#open-weight-share-by-model-type)
- [US, Chinese, and Indian labs](https://theaiengineering.com/articles/open-vs-closed-ai-models#us-chinese-and-indian-labs)
- [Labs that released the most open-weight models](https://theaiengineering.com/articles/open-vs-closed-ai-models#labs-that-released-the-most-open-weight-models)
- [Indian labs in the tracker](https://theaiengineering.com/articles/open-vs-closed-ai-models#indian-labs-in-the-tracker)
- [The drop holds in every cut of the data](https://theaiengineering.com/articles/open-vs-closed-ai-models#the-drop-holds-in-every-cut-of-the-data)
- [Choosing between open-weight and proprietary models](https://theaiengineering.com/articles/open-vs-closed-ai-models#choosing-between-open-weight-and-proprietary-models)
- [Methodology](https://theaiengineering.com/articles/open-vs-closed-ai-models#methodology)
- [Download the data](https://theaiengineering.com/articles/open-vs-closed-ai-models#download-the-data)
- [Frequently asked questions](https://theaiengineering.com/articles/open-vs-closed-ai-models#faq)
