Granite 4.2: Reasoning, 512K Context, and Multi-Step RL

A new family of language models, Granite 4.2, focuses on a five-phase training process, a 512K token window, and tiered reinforcement learning. The 8B and 30B versions learn to manipulate tools in real-world environments, while a strict quality control framework surrounds a SFT corpus of 7.2 million examples. The entire set is published under the Apache 2.0 license and integrates with OpenAI-compatible endpoints.

An asynchronous, stepwise RL calibrated on batches of 4,096

After supervised fine-tuning, Granite 4.2 applies a reinforcement learning pipeline consisting of steps targeting distinct capabilities, from mathematical problem-solving to instruction following, then software engineering, terminal usage, and web searching. Each step corresponds to an independent pass that resumes from the checkpoint of the previous step. The training uses an asynchronous GRPO variant: workers continuously generate, a shared buffer collects trajectories, and the trainer optimizes without blocking generation. Policy refreshes can occur during the process, with workers reusing their KV cache. A limit prevents workers from falling more than one update behind the trainer, and the objective applies truncated importance sampling to cap the probability ratio and prevent a small number of outdated tokens from influencing an update. Advantages are calculated relative to the group with a leave-one-out base, eliminating the need for a separate value network. The first step, RLVR, pairs 256 prompts with 16 responses each, resulting in a batch of 4,096 examples consumed in a single optimization before the next pass. Subsequent steps maintain this mechanism. The 8B and 30B models also include an agentic RL block that teaches them to call tools, edit and execute code, use a terminal, and perform web searches in real-world environments.

SFT: OpenAI chat normalization, LLM judges, and SHA-256 deduplication

The SFT pipeline is preceded by a multi-step quality control process. Data from various sources is normalized into a unified OpenAI chat format, including structuring interactions with tools. Two models, GPT-OSS-120B and Gemma 4, evaluate the quality of the samples; those rated poorly, containing fabricated information or invalid tool interactions, as well as calls to unlisted functions, are removed. Targeted heuristics by dataset eliminate identified noise. Local and global deduplication, based on SHA-256 hashes calculated on message and tool fields, removes duplicates within each source and across the total mix.

A SFT corpus of 7.2 million samples, of which 31.6% are agentic

The SFT mix includes 31.6% agentic data and 68.4% non-agentic, totaling approximately 7.2 million samples and nearly 100 billion tokens, of which 65 billion are trainable. In the agentic category, software engineering accounts for 69%, tool calling 12.1%, terminal usage 8.0%, mathematics 3.5%, searching 0.8%, and actions 0.2%. These trajectories come from a wide range of harnesses and frameworks, including OpenHands, OpenCode, Terminus-2, SWE-agent, OpenResearcher, MiniSWE, OpenSeeker, EnvScaler, Gemini CLI, Hermes, Codex, and Goose, blending open-source games and internal synthetic RL environments. In the non-agentic category, instruction following and coding each total 18.8%, mathematics 14.6%, multilingual 7.0%, sciences 5.4%, reasoning 3.0%, and security 0.8%. This corpus aims to make the model a reliable assistant for instructions, reasoning, and tools.

Granite 4.2 uses a dense decoder transformer with RoPE

Granite 4.2 relies on a dense decoder-only transformer: GQA attention with 40 heads and 8 KV heads, RoPE positional encoding with θ=10,000,000, SwiGLU MLP, RMSNorm with ε=1e-5, unlinked input and output embeddings, and bfloat16 precision. The initial training of the models is conducted from scratch, on approximately 15T of tokens, spread across five distinct phases. The first two phases lay the groundwork, the third and fourth correspond to intermediate training marked by a gradual decrease in data quality, while the fifth phase allows for the integration of long context, extending the window to 512K tokens. At each phase, a specific data mix and a tailored learning rate schedule are used, evolving from vast web sources to more filtered corpora.

Integration: OpenAI endpoint, vLLM, and SGLang support

The models generate tool calls according to the OpenAI function call format and are compatible with endpoints such as vLLM, without requiring an additional integration layer for agentic harnesses. For interaction, each model includes a thought/non-thought switch, a lightweight reflection mode for simple situations, as well as the ability to generate a chain of thought before the response. Granite 4.2 is also supported in SGLang.

Range, Apache 2.0 license, and large-scale SFT tuning

The range covers three sizes, 3B, 8B, and 30B, sharing architecture and training pipeline, with capability gaps highlighted after post-training. The 30B receives a second SFT phase dedicated to agentic coding: oversampling of agentic data, SWE, and coding, retaining about 16% of repeat data, and an additional epoch with a learning rate of 3.0e-6. The final SFT executions rely on 32 to 128 nodes per model size, with each node offering 4 Grace/GB200 accelerators. The tuning includes a constant learning rate of 1.0e-5 after warmup (2.5% of iterations), TP=2, PP=1, and CP=4 or 2. The entire family is published under the Apache 2.0 license.