Pick Qwen3-8B on Hugging Face and download it. What lands on your disk is a folder: large files of weights, a configuration describing the network, a tokenizer, and a model card. This checkpoint’s license is Apache 2.0.1 Point a compatible open-source runtime at the folder, and a model built by Alibaba’s Qwen team can answer questions on your machine. You don’t need an API key, and no provider is metering each request.
That feels good to me, a lot like the move to Apache felt at TheStreet.com. But the folder contains a trained artifact. To understand which freedoms you have, it helps to separate that artifact from the process that made it and the service that might wrap it.
What a model is made of
GPT stands for generative pre-trained transformer.2 The transformer is an architecture: a design for how information moves through the network. A decoder language model such as Qwen3 breaks input text into tokens, converts them to numerical representations, and processes them through layers. Attention lets each position draw on earlier positions in its context. The model produces scores for possible next tokens; the runtime selects one, adds it to the sequence, and repeats.
The architecture describes the machine. The weights are its learned numerical parameters. In a downloaded checkpoint they appear as tensors, arrays of numbers, rather than a readable account of what the model knows. Some larger Qwen models use a mixture of experts, activating only part of their total parameters for a given token.3 Qwen3-8B is a smaller dense model, so the size and hardware demands of another checkpoint can be quite different.
The weights encode useful structure learned during training. They can be examined, changed, and tested. What they don’t provide is a record of how each training choice affected the result. Reading the numbers is not like reading a program’s source and finding the function responsible for a behavior.
For this book I separate nine layers of a deployed system. The first two are the architecture and weights. Then come the training data, the training procedure, and post-training. Inference code loads and runs the result. A service adds serving infrastructure, behavioral controls, and a deployment context that includes its API, geography, and data-retention rules.4 A release may expose some layers and leave others entirely with its publisher.
How the weights get made
The transformer architecture was introduced in the 2017 paper “Attention Is All You Need.” OpenAI’s GPT-2 followed in 2019 and GPT-3 in 2020.5 ChatGPT’s November 2022 launch brought language models to a much wider public.6 The familiar source-code vocabulary now had to describe an artifact produced through training.
Pretraining starts with data such as web crawls, books, code, licensed collections, and sometimes material generated by other models. Training adjusts parameters to improve an objective, commonly prediction of the next token. At large scales this can occupy clusters of accelerators for weeks or months. The source mix, filtering, duplicate removal, training schedule, hardware, and checkpoint selection all affect what is produced.
Post-training further shapes behavior. Supervised examples can teach an instruction format; reinforcement learning can reward preferred responses or successful task outcomes. Human feedback and AI feedback are both used.7 These stages can improve instruction following, tool use, reasoning, and refusals. They also introduce more data and procedural choices that an inference-only release may not disclose.
A useful technical report can describe these choices without supplying enough material to repeat them. Token counts and benchmark results tell a reader something. They don’t substitute for the data processing and training code needed to reconstruct the work.
Steering while the model runs
Behavior can change after training too. As a model processes text, its layers produce intermediate numerical states called activations. Researchers have identified directions in those activations associated with features such as tone or refusal. Adding or subtracting a direction during inference can change responses without changing the stored weights.8
Open weights let a researcher try such interventions directly. Fine-tuning offers another route, changing the parameters themselves. Both can produce meaningful modifications even when the original training corpus is unavailable. That is different from rebuilding the original training run.
A hosted service adds more possibilities. Its system instructions, filters, tool permissions, and serving settings can affect the answer a user receives. Access to one published checkpoint doesn’t establish which checkpoint a provider runs or which controls it applies around it. Evaluating the downloaded artifact and evaluating the service are separate jobs.
What the Qwen folder gives you
For Qwen3-8B, the public weights, architecture information, tokenizer, and compatible runtime give a user substantial control over inference. Apache 2.0 permits use, modification, and redistribution subject to its terms. That assessment belongs to this checkpoint and license, not automatically to every release bearing the Qwen name.
The Qwen3 report describes architecture, parameter and token counts, and training stages. It doesn’t publish the full corpus or the complete procedure needed to repeat training.9 A user can fine-tune the downloaded model or maintain a derivative. The release doesn’t give that user the same ability to inspect and change the upstream training process.
This is where separate questions help. Can I run the artifact without a provider? Can I modify and redistribute it? Can I trace the training sources and transformations? Can another group build a substantially equivalent model from the published materials? A yes to the first two doesn’t answer the last two, and a missing recipe doesn’t erase the useful control the first two provide.
What OSI asks for
The Open Source Initiative’s AI definition requires the preferred form for making modifications: parameters, the complete code used to train and run the system, and sufficiently detailed information about the training data. That information must let a skilled person build a substantially equivalent system.10 It is more demanding than an inference-only release under a permissive license.
It does not require every training datum to be publicly redistributed. The definition explicitly accommodates unshareable data, while requiring descriptions of its provenance, characteristics, selection, labeling, and processing. Publicly available data and data obtainable from third parties must be identified. Whether a particular disclosure is sufficient remains a substantive question, rather than a box marked “data mentioned.”
I don’t think publishing an entire large training corpus is always pragmatic, especially where the data can’t legally be shared. That concern alone isn’t a disagreement with OSI. The practical dispute is how much information and code a release supplies, whether someone else can use it to reproduce the work, and which freedoms remain possible without it. Compute and funding are further limits: having permission and a recipe doesn’t pay for the training run.
My own minimum for useful control is still the ability to run and modify a model on hardware I choose. I would also want publishers to explain their training sources and methods. Those are practical preferences, not a substitute definition of open source. The next chapter uses Llama’s published card and terms to show how to read the evidence for each kind of claim.
Notes
Qwen Team, “Qwen3: Think Deeper, Act Faster,” Qwen blog, 29 April 2025, https://qwenlm.github.io/blog/qwen3 (the Qwen3 dense and MoE models are released “under Apache 2.0 license”); Qwen, Qwen3-8B LICENSE (Apache License, Version 2.0), Hugging Face, https://huggingface.co/Qwen/Qwen3-8B/blob/main/LICENSE. ↩
OpenAI, “Key concepts,” OpenAI API documentation, https://developers.openai.com/api/docs/concepts (OpenAI’s text generation models are “often referred to as generative pre-trained transformers or ‘GPT’ models for short”), accessed 15 November 2026. The approach comes from Alec Radford, Karthik Narasimhan, Tim Salimans, and Ilya Sutskever, “Improving Language Understanding by Generative Pre-Training,” OpenAI, 2018, https://cdn.openai.com/research-covers/language-unsupervised/language_understanding_paper.pdf, which pre-trains a transformer as a generative language model and then fine-tunes it. ↩
Noam Shazeer et al., “Outrageously Large Neural Networks: The Sparsely-Gated Mixture-of-Experts Layer,” arXiv:1701.06538, submitted 23 January 2017, https://arxiv.org/abs/1701.06538. For Qwen, the Qwen3 announcement lists Qwen3-235B-A22B with 235 billion total and 22 billion activated parameters: Qwen Team, “Qwen3: Think Deeper, Act Faster,” 29 April 2025, https://qwenlm.github.io/blog/qwen3. ↩
The nine-layer breakdown is this book’s working framework, separating architecture, weights, training data, training procedure, post-training, inference code, serving infrastructure, behavioral controls, and deployment context. It distinguishes the downloaded model release from the service in which a model may be used. ↩
Ashish Vaswani et al., “Attention Is All You Need,” arXiv:1706.03762, submitted 12 June 2017, https://arxiv.org/abs/1706.03762; OpenAI, “Better language models and their implications,” OpenAI, 14 February 2019, https://openai.com/index/better-language-models/, with the GPT-2 paper, Alec Radford et al., “Language Models are Unsupervised Multitask Learners,” 2019, https://cdn.openai.com/better-language-models/language_models_are_unsupervised_multitask_learners.pdf; Tom B. Brown et al., “Language Models are Few-Shot Learners,” arXiv:2005.14165, submitted 28 May 2020, https://arxiv.org/abs/2005.14165. ↩
OpenAI, “Introducing ChatGPT,” OpenAI, 30 November 2022, https://openai.com/index/chatgpt/. ↩
Long Ouyang et al., “Training language models to follow instructions with human feedback,” arXiv:2203.02155, submitted 4 March 2022, https://arxiv.org/abs/2203.02155 (human labelers rank model outputs to train a reward model); Yuntao Bai et al., “Constitutional AI: Harmlessness from AI Feedback,” arXiv:2212.08073, submitted 15 December 2022, https://arxiv.org/abs/2212.08073 (reinforcement learning from AI feedback). ↩
Alexander Matt Turner et al., “Steering Language Models With Activation Engineering,” arXiv:2308.10248, submitted 20 August 2023, https://arxiv.org/abs/2308.10248 (the activation-addition method); for a refusal direction specifically, Andy Arditi et al., “Refusal in Language Models Is Mediated by a Single Direction,” NeurIPS 2024, arXiv:2406.11717, https://arxiv.org/abs/2406.11717. ↩
Qwen Team, “Qwen3: Think Deeper, Act Faster,” Qwen blog, 29 April 2025, https://qwenlm.github.io/blog/qwen3; An Yang et al., “Qwen3 Technical Report,” arXiv:2505.09388, submitted 14 May 2025, https://arxiv.org/abs/2505.09388. ↩
Open Source Initiative, “The Open Source AI Definition 1.0,” https://opensource.org/ai/open-source-ai-definition. Its preferred form for modifications requires data information, complete training and running code, and parameters under appropriate terms. The data-information requirements explicitly include descriptions of unshareable data and listings of data available publicly or from third parties. See also OSI, “Open Weights,” https://opensource.org/ai/open-weights. ↩
From But Is It Open? by Timothy O’Brien (Discursive, 2026). © 2026 Timothy O’Brien. All rights reserved.