Ornith 1.5 vs Qwen vs Granite: 24GB context comparison

This local LLM comparison tests Ornith 1.5, Qwen 3.8, Qwen 3.6-A3B and IBM Granite 4.2 through Ollama at Q4_K_M quantization. The speaker uses a Radeon RX 7900 XTX with 24GB VRAM, a Ryzen 9 9950X and 64GB system memory. Each model runs at its largest context that fits entirely on the GPU, rather than at a shared context size.

The speaker's deterministic benchmark, Benchy, probes context capacity before running 21 tests. Reported ceilings are 128K for Ornith, 64K for Qwen 3.8, 32K for Qwen 3.6-A3B and 16K for Granite. Ornith passes all 21 tests with a 99.3 overall score; the others score 98.7, 90.2 and 86.5 respectively. Qwen 3.6-A3B generates fastest at 135 tokens per second, but trails Ornith in coding and reasoning despite their similar mixture-of-experts configurations.

Granite illustrates two separate runtime costs. In the speaker's check, disabling thinking reduces one response from 269 to 91 tokens without materially changing generation speed. Raising its context to 128K requires 53GB, spills into system memory and drops generation to 5.4 tokens per second.

These are results from one hardware configuration and benchmark suite. The speaker does not isolate each model's attention design or test a 32GB card or Mac Studio. Benchy is described as a private repository shared with the Dynamous community.