開放模型
IBM Granite 4.2 Opens Three Dense Reasoning Models, but Official Documents Disagree on Their Training Starting Point
Granite 4.2 introduces 3B, 8B, and 30B models with switchable thinking modes, native tool calling, and context windows of up to 512K, under the Apache 2.0 license. The 8B and 30B models underwent sandbox-based agentic reinforcement learning, but the model card and technical article contradict each other on whether the models were pretrained from scratch.
IBM released Granite 4.2 on August 25, offering three decoder-only dense Transformers in 3B, 8B, and 30B sizes. All three support full thinking, low-compute thinking, and no-thinking modes, and can produce native function calls through OpenAI-compatible endpoints in vLLM and SGLang. The weights are licensed under Apache 2.0. The models have a native sequence length of 128K, while IBM says additional long-context training extends support to 512K. [IBM Research](https://research.ibm.com/blog/introducing-granite-4-2) positions them as a family of agentic models that can be deployed across cloud, on-premises, and edge environments.
The technical advances extend beyond the addition of `<think>` tags. IBM’s published training description says the SFT dataset contained approximately 7.2 million examples totaling about 100 billion tokens, with agentic data accounting for 31.6%. The 8B and 30B models subsequently underwent additional agentic RL in real-world sandboxes, practicing terminal operations, code modification, search, and tool selection; the 3B model did not receive this training stage. IBM also released FP8, FP4, and multiple GGUF quantizations, lowering the barrier to self-hosted deployment. According to internal results listed in the [technical article](https://huggingface.co/blog/ibm-granite/granite-4-2), the 30B model scored 57.0% on SWE-bench Verified and 29.24% on Terminal-Bench 2.1, while the 8B model scored 47.67% and 20.56%, respectively.
However, engineering teams should first clarify a discrepancy concerning the models’ training provenance. The same technical article says all three models were “pretrained from scratch” on approximately 15 trillion tokens, while the 30B [model card](https://huggingface.co/ibm-granite/granite-4.2-30b) lists Granite 4.1-30B-Base as the base model and says Granite 4.2 was produced by post-training it. This affects assessments of data lineage, licensing, and reproducibility. The available agentic and long-context figures also come primarily from IBM’s internal evaluations. Although Chinese is listed among the tested languages, independent evaluations of Chinese tool calling are still lacking. The next points to watch are whether IBM corrects the documentation, publishes the complete evaluation setup, and discloses the KV-cache cost and accuracy degradation of real-world 512K deployments.