Inside a Language Model

The token counts are illustrative estimates for the visible text, not tokenizer measurements. System prompts, chat history, tools and hidden reasoning are excluded. The animated compute counter uses approximately 2 × 8.2 billion parameters per processed token, with one prefill and cached decoding; attention overhead is omitted. BF16 weight storage is 2 bytes per parameter; actual memory also includes activations, caches and runtime. These are Qwen3-8B scale calculations, not ChatGPT or Claude internals.

Environmental figures are Google’s reported median Gemini Apps text-prompt footprint from May 2025, not a prediction for this sentence or Qwen. They include serving infrastructure and data-center overhead; water refers to operational data-center cooling. They exclude model training, device use and other lifecycle impacts. The price is an illustrative energy-only conversion using €0.30/kWh, not a provider tariff or API price. Counters begin at submission and reach their reference totals by 59 seconds. Their interpolated progress is cinematic, not a measured resource-use curve. Weight memory is a fixed capacity, not a cumulative consumption.

• Google: energy, cooling water and emissions
• Qwen3-8B: official model card
• Research: approximate inference FLOPs

Sources checked 30 September 2026.