IBM just dropped a new family of open-weight models that quietly raises the bar for what enterprises can expect from freely licensed AI. The Granite 4.2 lineup spans 3B, 8B, and 30B parameter sizes, all trained from scratch on roughly 15 trillion tokens. That’s a serious training run, and it shows in the specs: context windows stretch to 512,000 tokens, which puts these models in direct competition with far more expensive proprietary systems when it comes to digesting long documents or entire codebases.
What makes this release genuinely interesting isn’t just scale — it’s the built-in agentic training. The 8B and 30B variants went through what IBM calls “agentic RL,” meaning they learned to use tools, write and execute code, and search the web inside real sandbox environments rather than simulated ones. IBM also baked in a thinking toggle: models can switch between deliberative and “low-effort” modes, letting developers dial compute up or down depending on whether a query is a simple lookup or a multi-step reasoning task. That kind of control is exactly what cost-conscious deployment teams have been asking for.
The speech side of the release is equally aggressive. Granite Speech 5.0 Turbo CTC models pack just 470 million parameters but claim to be twice as fast as previous leaders on the Open ASR Leaderboard. IBM says they can transcribe three hours of audio in one second. For call centers, meeting tools, or any real-time transcription workload, that’s the difference between batch processing and live feedback.
Everything ships under Apache 2.0 on Hugging Face, Ollama, GitHub, and other platforms, with support for OpenAI-format tool calling and deployment via vLLM or SGLang. The open-weight race has been heating up all year — Meta’s Llama, Mistral, and DeepSeek have all pushed hard on capability-per-parameter — but IBM’s focus on agentic behavior and a permissive license could make Granite 4.2 the default choice for companies that want tool-using models without vendor lock-in. Whether the real-world performance matches the spec sheet is the question that matters now.
