IBM releases Granite 4.2 at 3B, 8B, and 30B under Apache 2.0 with 512K context
IBM shipped Granite 4.2 in 3B, 8B, and 30B sizes under Apache 2.0, trained from scratch on about 15 trillion tokens with context windows to 512,000 tokens and a toggle between thinking and non-thinking modes. The 8B and 30B variants went through agentic reinforcement learning in sandboxes where they use tools, run code, and search the web, and all models support OpenAI-format tool calling on vLLM or SGLang. IBM published the full training recipe alongside the weights, so the agentic post-training method is reproducible rather than a black box — rare at this scale.
Source: the-decoder.com ↗
The 8B and 30B variants also go through what IBM calls "agentic RL" training, where they learn to use tools, write and run code, and search the web in real sandbox environments.
Why this matters
- → Open-weight agentic models with reproducible training—rare at 30B scale
- → 512K context and tool-use training enable complex autonomous workflows
- → Apache 2.0 license removes barriers to production deployment