Descriptions:
Fahd Mirza covers the release of IBM’s Granite 4.1 model family — available in 3B, 8B, and 30B parameter sizes — under a fully permissive Apache 2.0 license, meaning it can be used, modified, and shipped in commercial products without restriction. The video focuses on the 8B variant, which ships with a 131,072-token context window and multilingual support across 12 languages.
A significant portion of the video unpacks how Granite 4.1 was built. IBM used a five-phase staged training approach: starting with broad general web data, then sharpening progressively on math and code, then applying two rounds of increasingly curated data, and finishing with a dedicated long-context extension phase. Data quality was enforced through an LLM-as-judge system scoring every training sample across six dimensions — including correctness, completeness, and instruction-following — with hard rejection of anything flagged for hallucination. After training, the model went through four separate reinforcement learning refinement stages covering general capability, human preference alignment, knowledge calibration, and math reasoning.
Mirza then installs the model locally on Ubuntu using Ollama, running it on an Nvidia RTX 6000 with 48 GB of VRAM (the model loads in roughly 19 GB). The hands-on test tasks Granite 4.1 with generating a complete Python file-watcher CLI tool that auto-adds docstrings via AST parsing and a local Ollama API call — the model produces working, runnable code in a single pass.
📺 Source: Fahd Mirza · Published May 03, 2026
🏷️ Format: Hands On Build
