IBM just shipped Granite 4.2, a family of open reasoning models in 3B, 8B, and 30B sizes that breaks from the company’s earlier instruction-tuned assistants. Every model can emit a chain of thought before answering, and every model exposes a thinking / non-thinking switch plus a low-effort mode for easy questions. The 30B model posts a 57.00 on SWE-Bench Verified and a 29.24 on Terminal-Bench 2.1, which puts it in genuinely competitive territory for agentic coding work.

The training pipeline is where things get interesting. IBM pre-trained from scratch on roughly 15 trillion tokens, then ran a multi-stage reinforcement learning chain that includes an agentic RL block — but only on the 8B and 30B. The 3B gets foundational RL and alignment only. That single design choice explains most of the capability gap across sizes. During that agentic phase, models learned to edit code, drive a terminal, and run web searches inside real sandboxed environments. SFT used about 7.2 million samples, with software engineering making up 69% of the agentic slice.

All three models ship under Apache 2.0, which is a meaningful differentiator when Meta’s Llama models carry their own commercial restrictions and DeepSeek’s licensing remains murky. IBM also released GGUF quants down to Q4_K_M, so the 3B runs on a laptop through Ollama or LM Studio. The 30B targets enterprises with A100/H100-class capacity or FP8 serving on vLLM.

The company also dropped two Granite Speech 5.0 Turbo CTC models at 470M parameters. They ditch the LLM backbone entirely and use connectionist temporal classification, hitting an RTFx near 12,600 on a single H200 — roughly double what current speed leaders manage on the Open ASR leaderboard. That’s a transcription workhorse, not a conversational assistant.

What’s notable here is the explicit reasoning control. Most open models force you to choose between a thinking model and a non-thinking model. IBM baked the toggle into the chat template itself, which matters for production deployments where latency budgets differ per request. A low-effort mode that spends a short reasoning budget on easy questions is the kind of practical feature that gets overlooked in benchmark wars but saves real money at inference time.