# RTX 3090 runtime validation

**Result: PASS**

Validated the uploaded `Qwen3.8-27B-IQ1_M.gguf` on a Vast.ai NVIDIA GeForce RTX 3090 using Ollama 0.32.14 with the CUDA backend on 18 August 2026.

## Output

> The answer is 4, and the capital of France is Paris.

## Runtime evidence

- Model SHA-256: `131cdf5c1c4b547081543382b00434e9ebf3f8eb369ef3714550086074f80bdf`
- Quantization: `IQ1_M`; architecture: `qwen35`; parameters: 27.3B
- CUDA offload: 66/66 layers; llama.cpp CUDA model buffer: 6869.24 MiB
- Observed peak VRAM: 7772.0 MiB
- Observed peak GPU utilisation: 39.0%
- Observed peak power: 183.38 W
- Observed peak temperature: 53.0 C
- Deterministic generation: 24.06 tokens/s, normal stop

`REQUEST.json`, `RESPONSE.json`, `GPU_SAMPLES.csv`, `SERVER_CUDA_EXCERPT.log`, the complete Ollama server log, model hashes and the reproduction script are included in this directory.