IBM Releases Granite 4.2 Open-Weight Models With Agentic RL Training and 512K Context

IBM has released Granite 4.2, a family of open-weight language models available in 3B, 8B, and 30B parameter sizes, under the Apache 2.0 license. The models were trained from scratch on approximately 15 trillion tokens and support context windows up to 512,000 tokens, according to IBM.

Thinking modes and agentic training

Granite 4.2 models can toggle between "thinking" and "non-thinking" modes to control how much compute is spent per task. A separate "low-effort" mode is designed to conserve resources on simple queries, IBM says.

The 8B and 30B variants went through additional training IBM calls "agentic RL" — reinforcement learning where the models practice using tools, writing and executing code, and searching the web inside sandbox environments. All models support OpenAI-format tool calling and are compatible with vLLM and SGLang inference frameworks.

IBM published benchmark results showing the 30B model leading across three agentic and tool-use tests, though the company did not release the underlying numerical scores in the materials reviewed for this article.

Granite Speech 5.0 Turbo

Alongside the language models, IBM released Granite Speech 5.0 Turbo CTC, a speech recognition model with just 470 million parameters. IBM claims it runs twice as fast as the previous leaders on the Open ASR Leaderboard and can transcribe three hours of audio in one second.

Availability

All Granite 4.2 models are available under Apache 2.0 on Hugging Face, Ollama, GitHub, and other platforms. As open-weight releases, they can be self-hosted without per-token API fees; IBM has not disclosed managed-hosting pricing for the models through its own cloud offerings.

What this means

Granite 4.2 continues IBM's strategy of positioning smaller, permissively licensed models for enterprise deployment rather than competing directly on frontier scale. A 30B-parameter model with built-in tool-use training and a 512K context window targets a specific niche: organizations that want agentic capabilities — code execution, web search, tool calling — without sending data to a third-party API or paying per-token costs at scale.

The agentic RL training approach mirrors what OpenAI, Anthropic, and others have applied to their flagship models, but doing it at 8B and 30B scale under Apache 2.0 makes the resulting checkpoints directly usable in self-hosted pipelines. The Granite Speech model's claimed 470M-parameter footprint and real-time transcription speed also suggest IBM is optimizing for edge and cost-constrained deployment rather than raw benchmark leadership.

The benchmark claims — including the 30B model's lead on agentic tasks and the speech model's 2x speed claim — have not been independently verified and should be treated as vendor-reported until third-party evaluations are available.