IBM Granite 4.2 Open-Sources Three Reasoning Models: Real-Sandbox Reinforcement Learning Training, 30B Achieves 57% on SWE-Bench

On August 25, 2026, IBM released the Granite 4.2 series, consisting of three dense decoder language models: 3B, 8B, and 30B parameter sizes, all open-sourced under the Apache 2.0 license with support for download, fine-tuning, and commercial deployment without licensing negotiations. This marks the first time IBM has made "reasoning" capability a core design goal of the Granite series—each model natively supports chain-of-thought output and offers switchable thinking modes.

In terms of raw training scale, all three models were trained from scratch on approximately 15 trillion tokens, using a five-stage progressive training strategy to extend the context window from the standard 128K to 512K tokens. The supervised fine-tuning phase covers approximately 7.2 million samples, including chain-of-thought reasoning, instruction following, and agentic trajectory data, of which agentic samples account for 31.6% and non-agentic samples 68.4%. Among the agentic samples, software engineering tasks further account for 69%—a figure that directly illustrates IBM's priority ordering.

Core of the Training Mechanism: Real Environments, Not Simulations

The most noteworthy aspect of Granite 4.2 is its reinforcement learning pipeline design, particularly the Agentic RL stage exclusive to the 8B and 30B models.

IBM adopted the asynchronous GRPO (Group Relative Policy Optimization) algorithm, which eliminates the separate value network compared to traditional PPO, streamlining the training process. The entire RL process is divided into multiple independent stages, each focused on a single capability and hot-started from the checkpoint of the previous stage.

The key difference lies in the authenticity of the training environment. The Agentic RL stage encompasses three scenarios: the first is software engineering—the model directly edits real code repositories through the OpenHands toolchain, with reward signals derived from whether hidden test suites pass. There is no simulated scoring; only whether the code runs or not. The second is terminal operation—the model executes tasks in an active Linux Shell environment, with up to 64 rounds of environment interaction allowed per inference. The third is web search—the model answers multi-hop questions through real-time search calls, with another large model serving as judge for scoring.

This design path clearly diverges from current mainstream industry practices. According to IBM's official blog, many open-source agentic models rely on synthetic trajectories for post-training—that is, "fake execution" data generated by other models. Granite 4.2's 8B and 30B chose the more costly and engineering-intensive path: letting the model learn to execute directly in a real sandbox, with reward signals determined by task outcomes rather than model judgment.

In terms of training infrastructure, IBM completed training on CoreWeave-hosted NVIDIA GB200 NVL72 clusters, with a single NVLink domain containing 72 GPUs and nodes interconnected via 400 Gb/s InfiniBand.

Benchmark Numbers and What They Actually Mean

IBM published benchmark results across multiple dimensions. On SWE-Bench Verified (a benchmark measuring a model's ability to genuinely fix GitHub issues), the 30B model achieved 57.00%, and the 8B model 47.67%. On the terminal operation benchmark Terminal-Bench 2.1, the 30B scored 29.24% and the 8B 20.56%. In mathematical reasoning, the 30B reached 89.17% on AIME25, the 8B 86.67%, and the 3B also reached 78.33%. On the scientific reasoning benchmark GPQA, the 30B scored 66.41% and the 8B 64.14%. On the long-context benchmark RULER (128K), the 30B scored 81.38%, the 8B 71.41%, and the 3B 55.30%.

The 57% SWE-Bench Verified score needs to be understood in the correct frame of reference. The significance of this number lies not in horizontal competition with frontier closed-source models—top closed-source models currently score significantly higher on this benchmark—but rather in the fact that it is achieved with 30B parameters, under the Apache 2.0 license, in a locally deployable form. For enterprises that require full control over the reasoning pipeline, this combination of parameter size and licensing terms previously had no comparable counterpart in the open-source ecosystem offering similar agentic capabilities.

IBM also released three reasoning modes: a default full chain-of-thought mode (outputting the complete reasoning process in dedicated tags), a non-thinking mode with direct answers, and a low-spending mode in between (consuming a shorter reasoning budget on simple problems). In multi-turn conversations, reasoning from historical turns is trimmed by default to conserve context. Tool calls use the OpenAI function calling format and can be directly integrated into agentic frameworks deployed with vLLM or SGLang without additional adaptation layers.

What It Means for Different Stakeholders

For individual developers and small teams, the 3B model already has GGUF quantized versions (down to Q4_K_M) that can run on laptops via Ollama or LM Studio, lowering the hardware barrier for agentic tool-calling experiments. The 8B model is well-suited for medium-sized teams with a single modern GPU.

For enterprise users, particularly in regulated industries such as finance, healthcare, and government, Apache 2.0 combined with locally deployable weights provides a path to run tool-calling agents without sending data to third-party APIs. This combination previously lacked high-quality agentic model support in the open-source ecosystem, and IBM's release fills that gap. The 30B model requires A100 or H100-class GPUs, or can be deployed with FP8/NVFP4 quantization on vLLM.

For agentic framework developers, Granite 4.2's compatibility design lowers integration costs: OpenAI-format tool calling means that frameworks already supporting mainstream API formats can integrate without modification. This is a direct benefit for ecosystems such as LangChain, AutoGen, and OpenHands.

Direct competitive pressure on IBM's rivals is relatively limited but not negligible. Granite 4.2 is not challenging GPT-5.5 or Claude Opus 4.8 on single-turn answer quality; rather, it establishes a benchmark position in the niche track of "agentic models enterprises are willing to deploy locally." The competitive logic of this track differs from general capability rankings: licensing terms, deployment costs, and compliance often carry more weight than absolute benchmark scores in purchasing decisions.

The Significance of the Training Approach as a Precedent

From a broader perspective, the Granite 4.2 release establishes a concrete technical reference point for open-source agentic training. Over the past year, industry discussions about "whether models can truly execute tasks in real environments" have largely remained at the demonstration level, whereas IBM provides a traceable training design: using OpenHands as the software engineering training harness, with success determined by hidden test suites; using real shells with 64 rounds of interaction; and using real web search to complete multi-hop reasoning. The publication of these design details in itself offers a reproducible path for the open-source community.

Also released alongside Granite 4.2 is the Granite Speech 5.0 Turbo CTC series, two 470M-parameter speech models that can transcribe 3 hours of audio in approximately 1 second on a single H200 GPU, claiming roughly twice the speed of the previous leader on the Hugging Face Open ASR leaderboard.