About

QwenNotes is a practitioner's notebook on the open Qwen family. We do not rehash press releases and we do not publish leaderboards: we write about what we ran ourselves, and about where a model fell over.

It began with one task — pulling line items out of several thousand scans on a machine that could not be connected to external services. It turned out there was plenty of material on "which model is better in general" and none on "what to do with this particular contract".

How we work

We verify on our own data. Every number in these notes came from our document set; the conditions are described next to it.
We write about limits. The "what not to expect" section matters more to us than the achievements section.
We do not pass opinions off as facts. An opinion is labelled as an opinion, a measurement as a measurement with conditions.
We check specs against the source. Prices and service limits change more often than our notes are published.

A practical note is useful if it lets you decide today, not after you have bought a cluster.

Method

What we measure and how

Measurements are taken on one machine, under identical conditions and with generation constrained by a grammar. Anything that cannot be reproduced twice does not make it into the notes.

Parameter	Value	Comment
CPU	8 cores x86-64	No accelerator: if it works here, it works on a server
Memory	32 GB	Enough for 8B at int4 together with context
Accelerator	one card	Separate runs for 27B: it works on the CPU, but slowly
Runtime	llama.cpp	Plus a server runtime for batch runs
Quantisation	Q4_K_M	The memory/quality compromise; Q5/Q6 get their own runs
Temperature	0	Extraction needs reproducibility, not variety
Output constraint	schema grammar	The model physically cannot emit invalid JSON
Dataset	400 documents, 4 types	Contracts, invoices, forms, statements; Russian language
Metrics	JSON on the first try, per-field accuracy	One "bad" field usually explains the whole difference

The configuration changes when more interesting hardware shows up. The current conditions are always stated in the note that contains the measurements.

What you will not find here

"Top 10 models" leaderboards. A model without a task is just a parameter count;
Weight download links. We link to the official model cards and never re-host files;
Paid reviews. These notes exist because we need them in our own work;
On-demand support. Notes are updated as time allows, not on request.

When we get a number or a condition wrong, we correct it and note separately what changed.

Questions

Short answers about the site

No. QwenNotes is an independent community site. We do not publish weights, we do not hand out access, we do not advise on licences and we are not responsible for how the models behave.

Yes, with a link to the source. Better still, reproduce them on your own documents: the gap between domains is usually larger than the gap between the models we compare.

A model without a task is just a parameter count. We describe the conditions and the result on a specific document set, and leave the conclusion to the reader.

We read material like this in the evening, with long tables. A dark theme with monospaced figures and a hairline grid is less tiring when working through measurements.