gemma3:4b
An open-source vision-language model in Google's Gemma 3 series with 4B parameters.
Timeline (last 90 days)
In the small sample of the 14-clause evaluation set, gemma3:4b made no errors of misjudging a binding clause as harmless (yes→no was 0).
Moonlight's tech stack uses Node.js, Express and Ollama, runs the gemma3:4b model, and has only one application dependency.
Moonlight's first version had gemma3:4b classify each clause into three categories (yes/unclear/no), and all 8 clauses tested returned unclear; after rewriting the prompt and adding two examples, the same model returned yes for all of the same 8 clauses.
On the 14-clause evaluation set, gemma3:4b achieved 50% classification accuracy (13 of 14 clauses were judged yes) and 100% verbatim excerpt accuracy (all 14 quotes matched verbatim).
The gemma3:1b model is 815MB in size, gemma3:4b is 3.3GB, and gemma3:1b is 15% faster.
On a GPU-less Linux machine (8 vCPU), gemma3:1b takes about 6 to 9 seconds per analysis but often classifies all conversations as asking_questions and gives generic replies; gemma3:4b takes about 15 to 24 seconds per analysis, can select a reasonable stage and reuse facts the seller mentioned, and can reply in Chinese.
After keyword hints were introduced, the 1B model went from always answering asking_questions to usually selecting the correct stage, while the 4B model sometimes disagreed with the hints, which the author believes indicates the model is making its own judgment.
In testing, the 4B model sometimes claimed it had already performed some action, added comparisons the seller had never mentioned, and in Chinese even got family relationships wrong (writing "your dad" for "the child's dad").
TerraGemma calls gemma3:4b for deeper reasoning on complex wilderness first aid and plant identification.
All Facts (9)
In the small sample of the 14-clause evaluation set, gemma3:4b made no errors of misjudging a binding clause as harmless (yes→no was 0).
50%UnverifiedMoonlight's tech stack uses Node.js, Express and Ollama, runs the gemma3:4b model, and has only one application dependency.
50%UnverifiedOn the 14-clause evaluation set, gemma3:4b achieved 50% classification accuracy (13 of 14 clauses were judged yes) and 100% verbatim excerpt accuracy (all 14 quotes matched verbatim).
50%UnverifiedThe gemma3:1b model is 815MB in size, gemma3:4b is 3.3GB, and gemma3:1b is 15% faster.
50%UnverifiedOn a GPU-less Linux machine (8 vCPU), gemma3:1b takes about 6 to 9 seconds per analysis but often classifies all conversations as asking_questions and gives generic replies; gemma3:4b takes about 15 to 24 seconds per analysis, can select a reasonable stage and reuse facts the seller mentioned, and can reply in Chinese.
50%UnverifiedAfter keyword hints were introduced, the 1B model went from always answering asking_questions to usually selecting the correct stage, while the 4B model sometimes disagreed with the hints, which the author believes indicates the model is making its own judgment.
50%UnverifiedIn testing, the 4B model sometimes claimed it had already performed some action, added comparisons the seller had never mentioned, and in Chinese even got family relationships wrong (writing "your dad" for "the child's dad").
50%UnverifiedTerraGemma calls gemma3:4b for deeper reasoning on complex wilderness first aid and plant identification.
50%UnverifiedMoonlight's first version had gemma3:4b classify each clause into three categories (yes/unclear/no), and all 8 clauses tested returned unclear; after rewriting the prompt and adding two examples, the same model returned yes for all of the same 8 clauses.
50%