IBM Releases Granite 4.2 Reasoning Models and Granite Speech 5.0

IBM expands Granite with reasoning, agentic RL and compact speech models

IBM released Granite 4.2 on August 25, 2026, adding explicit reasoning and stronger agentic training to its open enterprise model family. The release includes dense decoder-only language models at 3B, 8B and 30B parameters, all under the Apache 2.0 license.

IBM also released two 470M-parameter Granite Speech 5.0 Turbo CTC automatic-speech-recognition models. The commercial-use model is Apache 2.0 licensed, while the separate NC variant uses a non-commercial CC-BY-NC-SA-4.0 license.

These are released models available now, not roadmap previews.

What is new in Granite 4.2

IBM says Granite 4.2 is its first dense Granite family with native reasoning. Every size can operate in thinking, non-thinking or low-effort thinking modes and supports native tool calling.

The 8B and 30B models receive an additional agentic reinforcement-learning stage in which they interact with real environments for software engineering, terminal operation and web search. The 3B model receives foundational RL and alignment but does not go through that full agentic block.

IBM's technical report says the models were pretrained from scratch on roughly 15 trillion tokens and use a five-phase strategy that eventually extends context training to 512K tokens. Supervised fine-tuning mixes reasoning, instruction, coding and agentic trajectories before staged RL and RLHF.

Training agents in real environments

The agentic RL sequence for the 8B and 30B models progresses through software-engineering, terminal and search environments.

For software engineering, models work inside repository sandboxes and are rewarded when hidden tests pass. Terminal training uses a live shell and multi-turn command execution. Search training uses web-search tools and an LLM judge for open-ended answers.

IBM reports benchmark results across SWE-Bench, Terminal-Bench, tool use, reasoning, instruction following and long-context tests. Those figures are IBM-reported evaluation results and should be interpreted in the context of the published methodology rather than as universal guarantees of production performance.

Granite Speech 5.0 takes a different architecture path

The two Granite Speech 5.0 Turbo CTC models are compact encoder-only ASR systems rather than speech models built around an LLM backbone.

IBM says the commercial model was trained on about 60,000 hours of English audio and uses 16 Conformer blocks, block attention, self-conditioning and temporal subsampling. In IBM's official H200 testing, throughput exceeded 12,600 RTFx, meaning batched inference can process hours of audio per second under that benchmark setup.

The NC variant uses additional restricted-use training data and is slightly more accurate on several reported tests, but its license restricts it to non-commercial use.

Why the release matters

Granite 4.2 combines open licensing with explicit reasoning, native tool calling and agentic post-training at sizes that can be deployed across cloud, on-premises and smaller infrastructure.

Granite Speech 5.0 targets a different bottleneck: high-throughput and edge-friendly transcription without an autoregressive decoder.

The release therefore covers two distinct enterprise-AI needs—models that can reason and act with tools, and compact speech recognition that can process large audio workloads efficiently.