Team AstraX evaluated five open-weight AI models: GPT-OSS-20B, GPT-OSS-120B, Qwen 3.8 27B, Gemma 4 31B, and Gemma 4 26B A4B.

We started by recording each model’s specifications from its official model card, including parameters, architecture, context length, licence, and intended use. We then designed a fair test with five categories: reasoning and mathematics, factual accuracy, instruction following, Indian languages, and safety and ethics.

Every model received the same prompts and settings, with a new conversation for each prompt. Two team members scored each response independently and agreed on the final score. Testing was done in the free Groq Playground and Google AI Studio.

The project includes a scorecard spreadsheet with the rubric and per-model and per-category results, and a presentation covering strengths, weaknesses, failure patterns, and recommendations. The main finding is that no single model is best for every task, so the right choice depends on the application.