Three new open-weight models bring native chain-of-thought reasoning and a 128K context window to self-hosted enterprise AI.
IBM released its Granite 4.2 family of large language models on August 25, a set of open-weight, locally deployable models. Three variants ship at once: a compact 3B model aimed at edge devices, an 8B mid-range option, and a 30B heavyweight that can stretch its context window to 512K tokens.
All three are released under the Apache 2.0 license, meaning companies can download, modify, and deploy them without royalty obligations or proprietary restrictions.
What Granite 4.2 actually does
The headline architectural feature is a native 128,000-token context window across the entire lineup. IBM built these as decoder-only models and layered in switchable chain-of-thought reasoning modes that let the model toggle between fast responses and deeper step-by-step inference depending on the task.
The 8B and 30B variants received additional treatment through an agentic reinforcement learning block, a training process that ran the models through simulated real-world environments to teach them how to operate terminals, execute code, search the web, and interact with external software tools. The 3B model supports tool-calling too, but without that specialized agentic training phase behind it.
All three were post-trained on approximately 15 trillion tokens, combining synthetic code generation with multi-stage reinforcement learning techniques.
How the benchmarks look
IBM published scores across two well-known benchmarks. On AIME25, which tests mathematical reasoning, the 3B model scored 78.33 and the 30B scored 89.17.
On SWE-Bench Verified, which measures a model’s ability to resolve real GitHub software engineering issues autonomously, the 8B scored 47.67 and the 30B scored 57.00.
IBM also launched Granite Speech 5.0 Turbo CTC models alongside the language release, expanding the Granite family into audio and speech processing.
Why the local deployment angle matters
The models are available through Hugging Face, Ollama, and GitHub. They also support OpenAI-compatible tool calling and inference frameworks like vLLM and SGLang, meaning teams can swap Granite models into existing agentic pipelines without rewriting the integration layer.
The 3B variant is specifically designed for edge and resource-constrained environments, the kind of hardware that lives in factory floors, retail endpoints, or field devices where you cannot run a 70B parameter model and cannot guarantee an internet connection.
Granite 4.1 shipped in April 2026, making this roughly a four-month iteration cycle.
The 30B model’s extendable 512K context window is the capability most likely to drive adoption in specific high-value enterprise verticals: legal document review, financial filings analysis, and large codebase comprehension are all tasks where context length directly determines whether a model is useful.
