IBM has released Granite 4.2, a family of open reasoning language models available in 3B, 8B, and 30B parameter sizes. Unlike previous iterations designed as instruction-following assistants, this release prioritises explicit reasoning. Every model can emit a chain of thought before providing an answer. Users can switch between thinking, non-thinking, and low-effort modes, which allocates a shorter reasoning budget for simple queries. The models are decoder-only dense transformers pre-trained from scratch on roughly 15 trillion tokens. They undergo a multi-stage reinforcement learning chain. The 8B and 30B versions include an agentic RL block where the system learns to edit code, drive a terminal, and run web searches inside sandboxed environments. All three models ship under the Apache 2.0 license. IBM also released two 470M-parameter Granite Speech 5.0 Turbo CTC models alongside the language models.
Deployment and licensing
All three Granite 4.2 language models ship under the Apache 2.0 license. This allows download, fine-tuning, and commercial production use without a licensing gate.
- Target users: The 3B model fits solo developers and startups running on a laptop via Ollama or LM Studio, especially with the released GGUF quants down to Q4_K_M. The 8B model suits mid-market teams on a single modern GPU. The 30B targets enterprises with A100/H100-class capacity or FP8/NVFP4 serving on vLLM. Regulated organisations benefit from the ability to host weights on-prem.
- Industries: Software and developer tooling, financial services, healthcare, telecom, public sector, and contact centers, which is where the new speech models are intended.
- Applications: Software engineering agents, terminal and DevOps automation, deep-research and search agents, long-document RAG, structured tool calling, and high-volume transcription.
Architecture details
Granite 4.2 is a decoder-only dense transformer, not a hybrid or MoE design. Core components include Grouped Query Attention with 8 KV heads, RoPE with θ = 10,000,000, SwiGLU MLPs, RMSNorm (ε = 1e-5), untied input/output embeddings, and bfloat16 precision.
The 3B model uses 40 layers at an embedding size of 2560. The 8B uses 40 layers at 4096. The 30B goes to 64 layers with an MLP hidden size of 32,768. The published architecture table lists a 131,072-token (128K) sequence length, while the five-phase pre-training run includes a long-context phase extending to 512K tokens. Pre-training covers roughly 15 trillion tokens from scratch.
The training pipeline
Supervised fine-tuning uses about 7.2 million samples, roughly 100B tokens with ~65B trainable. The mixture is 31.6% agentic and 68.4% non-agentic, and software engineering is 69% of the agentic slice. Trajectories were generated across harnesses including OpenHands, SWE-agent, Terminus-2, MiniSWE, Codex, and Goose. Quality control used GPT-OSS-120B and Gemma 4 as judges, plus SHA-256 deduplication over the tools and messages fields.
Post-training is a multi-stage, multi-environment RL chain, not a single pass. Each stage is a separate asynchronous GRPO run that warm-starts from the previous checkpoint, with a leave-one-out baseline instead of a value network and truncated importance sampling to bound off-policy drift. The order is RLVR, then skill boosters, then SWE, Terminal, Search, then RLHF.
The agentic RL block runs only on the 8B and 30B. The 3B takes foundational RL and alignment only. That single design choice explains most of the capability gap across sizes. Training ran on NeMo-RL and NeMo-Gym over an NVIDIA GB200 NVL72 cluster hosted by CoreWeave.
Two supporting pieces matter: 1 trillion tokens of synthetic code from IBM’s CodeAlchemy pipeline, and a speculative decoding layer for faster serving.
Reported results
IBM’s numbers, by size (3B / 8B / 30B):
| Benchmark | 3B | 8B | 30B |
|---|---|---|---|
| SWE-Bench Verified | NA | 47.67 | 57.00 |
| Terminal-Bench 2.1 | NA | 20.56 | 29.24 |
| τ³-bench | 50.99 | 66.34 | 68.05 |
| BFCL (v4) | 52.41 | 50.29 | 61.39 |
| AIME25 | 78.33 | 86.67 | 89.17 |
| GPQA | 54.80 | 64.14 | 66.41 |
| MMLU-Pro | 67.84 | 74.04 | 77.60 |
| RULER 128K | 55.30 | 71.41 | 81.38 |
Speech: 470M parameters, no LLM backbone
The Turbo CTC models come in at 470 million parameters and drop the LLM backbone entirely, using connectionist temporal classification to map audio to text. IBM reports an RTFx throughput near 12,600 on a single H200, against roughly 6,000 for current speed leaders on the Open ASR leaderboard. A WebGPU demo is live.
What it means
Developers now have a choice of models that can explicitly show their work, reducing hallucinations in complex tasks. The 3B model offers a lightweight option for local use without agentic capabilities, while the larger 8B and 30B versions provide the necessary training to handle code editing and terminal control. The open licensing ensures teams can deploy these models internally without legal hurdles. The new speech models offer a distinct path for transcription without relying on a large language model architecture.
