NutsNews 3B versus 14B summary screening

Historical screening results and their production limitations.

Visual overview

Primary diagram

System map

A small controlled model screening provides observations, but production stays unchanged until representative repeated chat tests and human review support a change.

Render the repository-owned system map when you need it.

Diagram is not rendered yet.

View as text

NutsNews 3B versus 14B summary screening decision

A small controlled model screening provides observations, but production stays unchanged until representative repeated chat tests and human review support a change.

flowchart TD
  accTitle: NutsNews 3B versus 14B summary screening decision
  accDescr: A small controlled model screening provides observations, but production stays unchanged until representative repeated chat tests and human review support a change.
  inputs["Three source-checked story digests"] --> test["One controlled run per story and model"]
  test --> results["Record speed, length and factual observations"]
  results --> limits["Small sample; not the production chat workflow"]
  limits --> decision["Keep production unchanged"]
  decision --> future["Repeated representative chat tests and human review"]

NutsNews 3B versus 14B summary screening

The September 13, 2026 screening compared Qwen 3B and 14B on the same three factual story digests. Each story was tested once with two inference threads. This was a controlled generation test, not the exact production chat workflow.

14B generated about 3.30 tokens per second versus 14.82 for 3B, roughly 4.5 times slower. Only one of its three summaries met the required length; 3B met it for all three. The larger model improved some wording, but the mouse-study summary used promotional language and did not explicitly say human lifespan benefits were unproven. Three examples do not establish general accuracy.

Keep production on 3B and improve the current validation pipeline first. Healthy web responses during the test do not establish unaffected AI latency. No publishing or application settings changed. The 14B model was unloaded, while its downloaded files were retained.