開放模型與代理訓練
Granite 4.2 Trains 8B/30B Models with Reinforcement Learning in Agentic Environments, Releases FP4, FP8, and GGUF Versions
IBM has released three Apache 2.0-licensed dense models in 3B, 8B, and 30B sizes, featuring switchable reasoning modes and native tool calling. The 8B and 30B models additionally underwent agentic reinforcement learning in codebase, terminal, and search sandboxes, although IBM’s official materials provide inconsistent accounts of their foundational training lineage.
IBM has released the Granite 4.2 family of 3B, 8B, and 30B dense decoder-only models, all under the Apache 2.0 license. All three support full thinking, non-thinking, and low-effort thinking modes, and can produce output in the OpenAI function-calling format. The 8B and 30B models additionally underwent agentic reinforcement learning, allowing them to edit and execute code, operate terminals, and search the web within realistic sandboxes.
The [technical overview](https://huggingface.co/blog/ibm-granite/granite-4-2) states that the architecture uses GQA, RoPE, SwiGLU, and RMSNorm, with a native context length of 131,072 tokens. The SFT corpus contains approximately 7.2 million examples and 100B tokens, with agentic data accounting for 31.6%; within the agentic portion, 69% focuses on software engineering. IBM used GPT-OSS-120B and Gemma 4 as data judges to filter out hallucinated content, invalid tool interactions, and calls to nonexistent functions.
The reinforcement learning pipeline uses NeMo-RL for training and NeMo-Gym for environments and verifiers, while generation and policy updates run asynchronously on separate GPU pools. IBM reports that the 30B model scored 57.00% on SWE-bench Verified, 33.29% on SWE-bench Pro, and 29.24% on Terminal-Bench 2.1, as well as 81.38 on RULER 128K. These results were measured by IBM and have not yet been independently ranked. For deployment, in addition to BF16 weights, IBM offers dynamic FP8, NVFP4/MXFP4 calibrated using 2,000 SFT samples, and multiple GGUF quantizations. The models can serve OpenAI-compatible endpoints through vLLM or SGLang.
One issue that warrants further investigation is source consistency: the technical article says the models were “pretrained from scratch” on approximately 15T tokens, while the [official GitHub repository](https://github.com/ibm-granite/granite-4.2-language-models) says Granite 4.2 was post-trained from the Granite 4.1 base model. IBM needs to clarify these conflicting descriptions of the models’ lineage. Engineering teams should also independently test the token savings of low-effort mode, tool-calling reliability after quantization, Chinese-language capabilities, and whether the models can maintain their reported sandbox performance in previously unseen agent frameworks.