Descriptions:
Fahd Mirza runs IBM’s newly released Granite 4.2 model family through a practical local benchmark, testing both the 3B and 8B parameter variants using vLLM on a 48 GB VRAM Ubuntu workstation. Granite 4.2 is fully open under Apache 2.0 and ships with native chain-of-thought reasoning, OpenAI-compatible tool calling, and a 128K context window — a notable feature set at the 3B scale. IBM has also published its complete training recipe: the 3B model receives reinforcement learning on math and code but skips the agentic training stages that the larger 8B and 30B variants undergo.
The evaluation task — generating a single HTML file listing every country per continent with each nation’s drink in native script — produces a clear split between the two models. The 3B fails outright: its tool call to save the output file never executes, the HTML structure is broken, and country coverage is severely incomplete with only four or five entries per continent. Switching to the 8B produces a working file with correct tab structure and substantially better coverage. VRAM consumption measures at roughly 4.5 GB for the 3B and over 45 GB for the 8B with a large KV cache configured. Mirza’s verdict is direct: the 3B is serviceable for straightforward reasoning, math, and code tasks but falls short for real agentic workloads, where the 8B or 30B variants are the appropriate choice.
📺 Source: Fahd Mirza · Published August 25, 2026
🏷️ Format: Benchmark Test
