IBM Granite 4.2: open 3B–30B reasoning models you can run locally
Apache 2.0 weights with a reasoning toggle, agentic RL that scores 57 on SWE-Bench Verified, GGUF quants for local runs, and new real-time speech models.
Copy markdown
Three open models, Apache 2.0
IBM shipped Granite 4.2 in 3B, 8B, and 30B sizes — all decoder-only dense transformers under a clean Apache 2.0 license, so you can download, fine-tune, and ship them commercially with no strings attached. Weights are live on Hugging Face with day-one vLLM and GGUF support.
A reasoning toggle you control
Every size ships with a thinking / non-thinking switch, plus a low-effort mode that spends only a small budget on easy prompts — so you pay for reasoning when it helps and skip it when it doesn't. The 30B posts 89.17 on AIME25 and 77.6 on MMLU-Pro.
Tuned to act like an agent
The 8B and 30B went through multi-environment agentic RL on software-engineering, terminal, and web-search tasks, plus 1T synthetic code tokens from IBM's CodeAlchemy pipeline. The 30B hits 57.0 on SWE-Bench Verified and 29.24 on Terminal-Bench 2.1 — enough to stand up a self-hosted coding agent this weekend.
It runs on your hardware
GGUF quants ship from Q8_0 down to Q2_K, so the 30B in 4-bit fits a single high-memory GPU or a 32GB+ Mac and the 3B runs on almost anything. Context stretches to 512K tokens for whole-repo or long-document work.
Elsewhere: open real-time speech
IBM also dropped Granite Speech 5.0 Turbo — 470M-parameter CTC models that hit roughly 12,600 RTFx on a single H200, fast enough for real-time transcription you host yourself.