IBM Granite 4.2 Brings Open Reasoning Models Local

A Reasoning Family You Can Actually Download

IBM released Granite 4.2 on August 25, 2026, and published the weights under Apache 2.0. That combination — a reasoning-oriented model family with a genuinely permissive license — is still rarer than the release cadence of the past few months might suggest, and it is the reason this release deserves a closer look than the benchmark table alone would earn it.

• Three sizes: 3B, 8B and 30B parameters, all released under Apache 2.0 along with quantized variants, the repository and the documentation
• Every model gets a switchable thinking mode, so the same weights serve both latency-sensitive calls and longer deliberate reasoning
• 512K context window, reached through a five-phase training schedule after pre-training from scratch on roughly 15 trillion tokens
• 57.00 on SWE-bench Verified for the 30B, with the 8B scoring 47.67 on the same benchmark

What the Switchable Thinking Mode Changes

Most deployments do not need extended reasoning on every request. Classifying a support ticket and refactoring a module are different jobs with different latency budgets, and running both through a model that always deliberates is an expensive way to be thorough.

Granite 4.2 exposes thinking as a toggle across all three sizes, including the 3B. That last part is the notable bit. Small models are usually shipped as the fast tier and left there; giving the 3B the same switch means an edge deployment can escalate its own reasoning depth without a round trip to a larger model on someone else's hardware. For anyone building on the kind of local model stacks we have been tracking, that is a meaningful architectural option rather than a spec-sheet flourish.

How Was Granite 4.2 Trained?

IBM describes pre-training from scratch on approximately 15 trillion tokens, followed by a five-phase schedule that progressively extends the context window out to 512K. The two larger sizes then went through reinforcement learning inside real environments — software engineering tasks, terminal sessions and web search — rather than on static preference data alone.

Training an agent in the environment it will actually work in is the more expensive path and generally the more honest one. Benchmarks like SWE-bench Verified reward exactly that kind of grounding, and the 30B's 57.00 is a respectable showing for a model sized to fit a single multi-GPU serving node.

Where the 30B and the 3B Each Fit

The 30B flagship is explicitly sized for one multi-GPU node, which puts it inside the budget of a well-equipped university lab or a mid-sized engineering team rather than requiring a rented cluster. The 3B targets lighter deployments — laptops, edge boxes, anywhere the memory ceiling is the binding constraint.

That spread matters for local LLM work more than a single headline number would. The open-weight releases of the past two weeks have mostly pushed upward in size: we covered GLM-5.3 Flash running locally on 128GB and Qwen3.8-Flash-Next fitting in 75GB without a GPU, and both are impressive, but neither is landing on a developer laptop. A 3B with a reasoning toggle and an Apache 2.0 license occupies a different and underserved slot.

The License Is the Quiet Headline

Apache 2.0 across the whole family — base models, quantized variants, code and documentation — means commercial use, modification and redistribution without a custom agreement to negotiate. For enterprises evaluating whether to build on open weights at all, license clarity routinely decides the question before capability does.

IBM has been consistent here across the Granite line, and consistency is its own feature. A team that standardized on Granite 4.0 knows what it is getting with 4.2, which is not something every open-weight publisher can claim right now.

Sources: IBM Granite via Hugging Face — August 25, 2026; MarkTechPost — August 25, 2026; Unite.AI — August 2026.