Built for speed
I built QNS-1 so summaries would feel instant on normal hardware, not like a heavyweight model pretending to be local-first.
ONCard in-house model
QNS-1
A lightweight summarizer model built for fast, local, factual summaries of study material, reports, articles, and technical text.
Parameters
400M
Size
367MB
Context
8K
Input
Text
Grounded summaries
The model is aimed at source-bound summarization: notes, reports, articles, webpages, and technical text where staying factual matters.
Structured output
It gives the quick answer first, then breaks the material into usable points so the result is easy to scan and study from.
English focused
For now I am keeping the scope narrow: English summarization, tuned carefully instead of spread thin across everything.
Run QNS-1 with Ollama.
Pull and run the public 400M model directly from Ollama for local summarization workflows.
Terminal
ollama run QyrouNnet/summarizer:400mBenchmarks
Benchmark results.
QNS-1 was evaluated against larger summarization models across overlap, semantic similarity, factual consistency, and reference-free evaluation metrics. In selected benchmark categories, the 400M model compares favorably with much larger systems, including models many times its parameter count.
| Metric | Score | Result | Measures |
|---|---|---|---|
| FactCC | 0.182 | 2nd | Factual consistency against the source text. |
| SummaC | 0.123 | 1st | Reference-free summary consistency. |
| ROUGE | 0.318 | Competitive | Lexical overlap with reference summaries. |