Granite 4.0 H Small has credible enterprise-oriented foundations: a 32B-total/9B-active hybrid MoE design, 131k context, instruction following, RAG, function calling, tool use, FIM coding, and multilingual dialog are documented by IBM. Its Apache-2.0 release and Hugging Face availability make it materially more deployable than closed API-only options. However, Artificial Analysis reports an Intelligence Index of 5, and the public Arena Hard result—rank 256, 1240±11 across 2,993 votes—does not support placing general capability or reasoning near Gemini 3.5 Flash, Kimi K2.5, GPT-5.5, or Claude Opus 4.8. Coding and agent-work scores remain conservative despite the documented tools and code features. DeepSWE and LiveCodeBench explicitly do not list the model; the supplied SWE-bench and Terminal-Bench/Aider materials provide no model-specific published result. This leaves it below DeepSeek-V3.2 and Kimi K2.5 on demonstrated agentic coding, while above Gemma 3’s low coding anchor only modestly on product support rather than benchmark proof. The model is text-centric in the supplied evidence, so its I/O score is closer to DeepSeek-V3.2 than multimodal Gemini or Claude offerings. Cost is the clearest advantage: IBM lists $0.0636/M input and $0.265/M output tokens, consistent with Artificial Analysis estimates, substantially under premium anchors. Artificial Analysis records high throughput (386.4 tok/s), but 10.22s TTFT and only one tracked provider limit the availability score. IBM’s documentation, watsonx API integration, open weights, and clear pricing support a strong developer score; limited independent quality evidence, a recent release, and weak preference standing constrain adoption and evidence-confidence scores.
