IBM releases Granite 4.1 Apache

IBM just made a very specific bet about what enterprise AI buyers actually want. Not the flashiest architecture. Not the biggest reasoning model. Something simpler — dense, open, inspectable, and easy to run in normal toolchains. That’s the point of Granite 4.1, which IBM released on April 29 as a broader family covering language, speech, vision, embeddings, and Guardian safety models, with the language side all under Apache 2.0. ### What actually shipped? The headline item is the language family: dense decoder-only models in 3B, 8B, and 30B parameter sizes, with both base and instruction-tuned variants. IBM also pushed out Granite Speech 4.1, vision, embeddings, and Guardian models in the same release, which makes this less like a single model launch and more like a full stack refresh for enterprise workflows. ### Why is “Apache” the big deal? Because IBM is not just saying “open-ish.” The Granite 4.1 language models are released under Apache 2.0, and IBM pairs that with cryptographic signatures, transparency disclosures, and ISO-related trust language in its docs. Basically, IBM is trying to remove the usual procurement headache around whether a company can actually adopt, modify, audit, and commercialize the weights without legal ambiguity. ### Why go back to dense models? This is the interesting turn. Granite 4.0 leaned into hybrid Mamba-2/transformer designs and used Mixture-of-Experts in some key models to cut memory use and speed inference. Granite 4.1 swings back to dense decoder-only models instead. IBM’s pitch is that the simpler architecture is more flexible for fine-tuning and easier to drop into the software stack companies already use. ### Is IBM giving up on MoE? Not exactly. Granite 4.0 is still there, and IBM’s own docs still frame those hybrid and MoE models as efficient for certain long-context and low-memory scenarios. But Granite 4.1 makes clear that IBM thinks a lot of practical enterprise work — tool calling, instruction following, coding, structured output — benefits from a more conventional dense setup, especially when depends. That’s an inference from how IBM is positioning the two families side by side. ### What’s the strongest performance claim? IBM’s sharpest claim is not “we beat everyone.” It’s narrower and more believable: the Granite 4.1 8B instruct model consistently matches or outperforms the older Granite 4.0 32B Mixture-of-Experts model on the tasks IBM cares about most here. Those tasks include instruction following and tool calling — exactly the boring-but-important stuff that makes enterprise assistants useful inside real software. ### How big is the context window? IBM says Granite 4.1 was trained through a five-phase process on about 15 trillion tokens, with a final long-context extension phase that scales the context window to 512K tokens. That is a big number, but the more practical point is that IBM wants these models to handle long documents, retrieval-heavy workflows, and codebases without needing a specialized architecture to do it. ### Who is this really for? Not consumers. Not frontier-model hobbyists chasing the absolute top benchmark. This is for teams that need to run models locally, wire them into tools, generate structured JSON, do RAG, and pass governance review without a month-long licensing argument. IBM even calls out edge deployment for the 3B model and general-purpose enterprise use for the 8B. Granite 4.1 is IBM saying the next enterprise AI fight may be less about exotic model design and more about trust, control, and operational friction. If an Apache-licensed 8B dense model can do the job of yesterday’s much larger MoE system, a lot of buyers will take the simpler option.