Archive: Three Qwen3 sizes, one useful local experiment
The original 8B, 14B and 32B dense releases make a controlled comparison easier. Start with one task you can grade.

In this article
From the release archive
On April 29, 2025, Qwen announced six dense open-weight models and two mixture-of-experts models. The dense 8B, 14B and 32B variants carry the Apache 2.0 license. The release offered both thinking and non-thinking modes, making the reasoning setting another variable to record.
Give each size the same job
Save a short support ticket and a checklist of facts a good summary must preserve. Run the same ticket through each model with the same prompt, reasoning mode and output limit. Keep a copy of the exact model revision and quantization for each run.
Read the result
Grade the summaries before looking at speed. Then ask whether a larger model fixed a meaningful omission and whether its extra wait is acceptable. The family announcement establishes what was released. Performance on your own ticket needs a local run.
Compare sizes using the same task and budget
Start with a task that has a checkable result, such as extracting five fields from a product spec into valid JSON. Run the same prompt and context against each size. Keep the same runtime and broadly similar weight precision. Record whether the JSON parsed and which fields were right. Then run a prompt close to the context length your application needs. A model that works on a short fixture may need a different quant or offloading arrangement when the input grows.
Record cold-load time separately from generation time. Keep track of CPU offloading and GPU memory. If a larger model is slow because part of it runs on the CPU, that result still matters to the buying decision for your machine. State the limitation alongside the result so other readers know what caused the difference.
What makes the archive useful
The three original dense variants give you a controlled set to learn how model size changes deployment on the same hardware. Use current candidates for a new purchase or project after you have a consistent measurement method.
Primary sources
Qwen release announcementRead next Hermes or Open WebUI? Choose by the job.