Qwen, an open-weight model, turned two Apple disclosure paragraphs into structured claims, quotations and a comparison. Explore the saved result, see what a small Python checker can verify, and decide which conclusions the evidence supports.
Start with the saved files. Neither route requires installing a model.
Why run your own model? Open weights give you choices about where AI runs, how you adapt it, and how you preserve an experiment. We use Qwen to explore that control through a task you can inspect. Explore the advantages or see why we chose this Qwen model.
A useful result
The task: compare two selected intellectual-property paragraphs from Apple's FY2024 and FY2025 filings. Here is one complete finding from the saved Qwen answer, arranged for reading. Its wording is unchanged.
Original Qwen output: Finding 2
Claim
Apple states that it is not always able to obtain all necessary licenses to third-party intellectual property on commercially reasonable terms or at all.
Passage ID P02 (Apple FY2025)
Type reported_statement
Exact quotation attached to the claim
The Company is not always able to obtain all necessary licenses to third-party intellectual property rights on commercially reasonable terms or at all.
That structure gives you something to work with: a claim you can inspect and a passage you can check. It reports Apple's stated licensing limitation; it does not establish what caused it. Read both selected source paragraphs.
What would you keep or change?
All five quotations in the saved answer match their sources. The first finding still needs a closer look. These short excerpts show the difference:
Source P02: Short excerpt
“may be exacerbated”
Original Qwen claim: Short excerpt
“exacerbates”
The source describes a possibility. The claim makes it sound definite. A matching quotation helps you find the evidence; it does not settle what you can say about it.
See the full claim and an editorial revision
Original Qwen finding 1: Apple discloses that the use of machine learning and artificial intelligence, including the training on copyrighted materials, exacerbates the risk of infringement allegations for its products and services.
Its exact P02 quotation: This risk may be exacerbated by the use of new and emerging technologies, including machine learning and artificial intelligence, which can involve, among other things, the acquisition and use of copyrighted materials for training as well as the potential reproduction of copyrighted materials in their outputs.
The claim drops “may,” even though its conditional_risk label and quotation preserve the idea of possibility. The source also says the technologies “can involve” copyrighted materials; it does not establish a particular training practice.
Assistant-proposed editorial revision: In the selected 2025 paragraph, Apple says that using emerging technologies such as machine learning and AI may exacerbate the risk of infringement allegations; such technologies can involve copyrighted training materials and potential reproduction of copyrighted materials in outputs.
This revision is separate from the original model answer. You can inspect both JSON files and the recorded changes in the download.
Go beyond one word
The comparison needs another kind of judgment. The selected 2025 paragraph reports licensing limitations and separately describes how AI may exacerbate infringement-allegation risk. Does that establish that AI caused the licensing difficulty?
Inspect the comparison and its proposed revision
Original Qwen comparison: Between the 2024 and 2025 filings, Apple's disclosure shifts from a generally optimistic view that licenses can be obtained on reasonable terms to an admission that it is 'not always able' to obtain them, and the 2025 filing explicitly links this risk to the use of machine learning and AI technologies, including the acquisition of copyrighted training data, a risk factor not present in the 2024 text.
Editorial review: P01 says licenses can generally be obtained, but immediately gives no assurance. P02 reports that necessary licenses are not always obtainable. Its AI sentence concerns infringement-allegation risk; these passages do not establish that AI caused licensing difficulty. The original comparison also omits the AI sentence from its own evidence.
Assistant-proposed editorial revision: The selected 2024 paragraph says licenses can generally be obtained on commercially reasonable terms but gives no assurance; the selected 2025 paragraph reports that Apple is not always able to obtain all necessary licenses on such terms or at all. The 2025 paragraph also adds a conditional AI/ML mechanism for infringement allegations. These wording changes do not establish that AI caused licensing difficulty or that the underlying risks first arose in 2025.
The revised JSON adds the exact AI sentence to the comparison evidence. The original answer and editorial revision both pass the checker; the decision about support remains a separate judgment.
Run the Python checks
Open the runnable example to check the saved Qwen answer with Python 3.9+ and its standard library. Keep the original answer, record one keep, revise or reject decision, and explain it with the source. Then inspect the separately labeled editorial revision. No API key or model download is needed.
For this saved case, stop when you have the check output and one supported editorial decision. Running another generation is optional; it is not needed to finish the exercise.
How the pieces work together
Define the output → ask the model → check format and quotations → review support → stop or revise.
| Piece | Its job in this example |
|---|---|
| Qwen | Proposes structured claims and attaches quotations |
| Python | Checks fields, passage IDs, exact quotation matches and both-year coverage |
| Your review | Checks claim strength, relationships and what the selected sources support |
| Stopping rule | Keeps the original visible and ends with a documented decision |
This is a fixed model-assisted workflow: its steps are predefined. The model does not choose its own tools or decide whether its answer is accepted. You are practicing structured extraction, output contracts, validation and evidence-based evaluation.
Why run an open-weight model?
Open-weight models make their learned parameters available to download. That opens up engineering choices:
- Control: Choose where the model runs and retain the version you tested.
- Customization: Adapt it to a specialized task through fine-tuning, within its license and your resources.
- Data privacy: Local inference can keep prompts and documents on your computer; connected services and the deployment setup determine what stays local.
- Potential cost savings: Repeated workloads may cost less, depending on hardware, utilization, energy and maintenance.
- Reproducibility: Preserve model files, code, prompts and settings to repeat an experiment. Identical outputs still depend on the runtime and hardware.
This Apple example starts with local execution and evaluation. A separate customization demonstration now explores adapter training; operating-cost comparisons remain future work. Technical sources and qualifications explain these possibilities.
Why this Qwen model?
Qwen3.8-27B gives us a local starting point with room to explore further. Its documentation describes tool use, multi-step planning and adjustable reasoning effort—features relevant to later agent experiments. Here, we use it to propose structured claims for a predefined checking workflow. We have not tested its autonomous tool use in this example. Official model documentation.
Our recorded setup: October 15, 2026
- Model
- Qwen3.8-27B, MLX 4-bit
- Reported model size
- 16.08 GB (not runtime memory)
- Computer
- Mac Studio, M4 Max, 36 GB memory
- Software
- LM Studio 0.4.26+4
- Loaded context
- 8,192 tokens
- Reasoning requested
- Off
What does 4-bit change—and how much memory is needed?
Quantization stores model weights using fewer bits per value. A simplified calculation for 27 billion parameters is 27 billion × 4 bits ÷ 8 = 13.5 GB. This estimates parameter storage, not the total memory needed to run the model.
LM Studio reported 16,081,678,492 bytes, or about 16.08 GB, for our model. Quantization metadata and components stored at higher precision can add to the simplified estimate. Running a model also needs memory for context and other runtime work. We did not measure peak memory, so neither number is a system-memory requirement. These GB figures use decimal units.
The setup above records the completed experiment, not a minimum hardware recommendation. The exact weights revision remains unknown; our saved settings and outputs support inspection but do not promise exact reproduction. The original model answer is available even if you cannot run this model yourself.
Did changing the weights help?
Classroom development note, October 16, 2026
We trained, saved and reloaded an adapter for a separate Qwen3-4B-Instruct-2507 model, using Python and MLX on the same Mac. An adapter adds a small set of trainable weights while keeping the base model fixed. This was not another run of the 27B Apple example above.
We used eight familiar claims from selected Notre Dame cleaned 10-Ks filed in 2021–2022: four training examples and four development probes. Ten training updates produced this comparison:
| Answers with an exact source quotation | Original instruction | Clearer instruction |
|---|---|---|
| Base model | 3/8 | 7/8 |
| Saved adapter | 3/8 | 7/8 |
Clearer instructions improved quotation capture; this tiny adapter added no improvement on these checks. Every condition returned valid JSON and matched all eight illustrative labels. Those labels were AI-proposed and source-checked, not independently validated answers; the familiar cases do not establish performance on new material.
One answer still quoted accurately while treating “certain businesses” as proof of “not every business.” The quotation matched, but the explanation went beyond what the words established. Another answer left its evidence field empty despite relevant wording.
We stopped there: customization worked technically, while its usefulness remained unproven. The download on this page covers the earlier Apple extraction example, not this adapter experiment.
In BUS243A, investigate when prompting, model adaptation and verification improve a business workflow—and when to stop.
Another worked run
The same download includes a separate native-assistant example with one generated correction, plus three deliberately constructed checks. Inspect that worked review to explore a different question: how narrowly should a claim be written to fit its assertion label? These artifacts are separate from the Qwen run and its editorial revision.
Source and execution notes for the Apple example retain the model settings, run history and artifact identities. The Qwen output is unchanged; the editorial revisions were prepared with AI assistance for your review.
Where this leads
A later investigation could let an agent retrieve missing context, route a difficult claim to another model, or request human help. Those extensions are planned. This example gives you the starting point: a model's output, checks you can run, and a decision you can explain.