Granite 3.3

tl;dr: IBM’s 128K-context text model pair, Apache 2.0, with reasoning and code-completion features dialed in. The 8B is the workstation sweet spot; the 2B is for laptops and embedded use.

What it is

granite3.3 is IBM’s text-only Granite release from April 2025. Two sizes — 2B and 8B, both dense, both 128K context, both with the same chat template. The 8B tag (granite3.3:8b) is the current default on the Ollama page at 4.9GB on disk; the 2B (granite3.3:2b) sits at 1.5GB.

The release notes call out gains on AlpacaEval-2.0 and Arena-Hard versus the 3.2 generation, plus improvements in mathematics, coding, and instruction following. The model supports Fill-in-the-Middle (FIM) for code completion tasks and exposes an explicit thinking mode.

Pull

ollama run granite3.3:8b   # default — 4.9GB on disk
ollama run granite3.3:2b   # 1.5GB on disk

What you get

• 128K context window on both sizes.
• Native function-calling support, listed as tools on the library page.
• Thinking mode, enabled by passing a control message with content: "thinking" before the user turn.
• Fill-in-the-Middle (FIM) for code completion.
• Multilingual out of the box: English, German, Spanish, French, Japanese, Portuguese, Arabic, Czech, Italian, Korean, Dutch, and Chinese.
• Listed capabilities include summarization, classification, extraction, QA, RAG, code tasks, function calling, multilingual dialog, FIM, and long-context tasks.
• Apache 2.0 license, declared on the library page.

Hardware

Gotchas

• granite3.3:latest resolves to the 8B tag — not the 2B. Pin granite3.3:2b if you want the small variant.
• Thinking output is verbose. Budget tokens for the reasoning trace.
• The “12 languages” list comes from the model card; users can fine-tune beyond those 12, but out-of-the-box support is the listed set.
• Apache 2.0 is clean — no Llama-style commercial restrictions — but downstream fine-tunes still need to respect the upstream terms.

Related: Granite 3.2 Vision for the multimodal counterpart, Ollama for the runtime, llama.cpp for the engine, Llama 3.1 for the Meta open-weights baseline at similar sizes.

Untested here — facts from ollama.com/library/granite3.3 fetched 2026-09-10.