IBM’s Granite 4.1 packs 30B dense model punch using just 15T tokens and a 5-stage data diet

IBM just dropped Granite 4.1, and the playbook they’re sharing is a direct challenge to the “scale is all you need” orthodoxy. This isn’t a mysterious black box; the team published the full recipe for their new family of dense, decoder-only LLMs, which come in 3B, 8B, and 30B sizes. The 8B instruct model is the immediate standout. It matches or beats the previous Granite 4.0-H-Small, a much larger 32B-A9B Mixture-of-Experts model. That’s a genuine efficiency gain, proving that a simpler dense architecture, when fed the right data, can absolutely embarrass a more complex one.

The secret isn’t just the 15 trillion tokens of training data. It’s the obsessive, five-phase pre-training pipeline that reads like a chef refining a broth. They start with a broad, web-scale mix heavy on CommonCrawl and code. Then they progressively anneal the data quality, boosting math and code to nearly 70% of the mix in Phase 2 before blending in chain-of-thought reasoning trajectories and synthetic instruction data. The result is a process that teaches the model fundamental language first, then forces it to learn structured reasoning, and finally polishes it with the highest-quality data as the learning rate decays to zero. It’s sequential learning, not a random jumble.

The context window story is equally pragmatic. A staged long-context extension process, complete with model merging after each jump to prevent short-context performance from crumbling, pushes the 8B and 30B models out to 128K tokens. The team then uses a specific diet of 80% books and 20% code repository data to hit that eye-popping 512K context ceiling. It’s a refreshingly specific and reproducible detail, not just a marketing bullet point.

What makes this release genuinely important is the open book. You can see the exact data mixture percentages shift across phases, the learning rate schedules, and the architectural choices like Grouped Query Attention. By also detailing their supervised fine-tuning with an LLM-as-Judge framework and a multi-stage reinforcement learning pipeline using on-policy GRPO, IBM is giving a masterclass in how to build a capable small model. The Apache 2.0 license is just the cherry on top.