IBM 詳細介紹如何建立開放 Granite 4.2 推理模型

發生了什麼事

IBM’s Granite team published a technical account of Granite 4.2, a new family of dense, decoder-only reasoning language models available in 3B, 8B and 30B sizes under the Apache 2.0 license. The models support thinking and non-thinking modes, low-effort reasoning, native tool calling and a context window extended to 512K tokens during training. IBM says the 8B and 30B versions were additionally trained to use tools, edit and run code, operate terminals and search the web inside sandboxed environments.

In a Hugging Face article published August 25, 2026, IBM’s Granite team described Granite 4.2 as the first Granite family built as dense, decoder-only reasoning models. The family has three sizes: 3B, 8B and 30B parameters. All use the same broad architecture, including grouped-query attention, rotary position embeddings, SwiGLU feed-forward layers, RMSNorm and bfloat16 precision. The article says the models were pretrained from scratch on approximately 15 trillion tokens through five phases, with the final phase extending the context window to 512K tokens. The architecture table lists a 131,072-token sequence length for the models, while the training strategy is described as extending context to 512K; the source does not explain that distinction in detail.

The post-training recipe is the central change IBM describes. Supervised fine-tuning used about 7.2 million samples, or roughly 100 billion tokens, combining agentic and non-agentic material. IBM says the data included software engineering, tool calling, terminal use, search, mathematics, multilingual instruction following, science, reasoning and safety examples. The company says it normalized the data into a common chat format, used GPT-OSS-120B and Gemma 4 as language-model judges, removed low-quality or invalid examples, and applied heuristic filtering and SHA-256-based deduplication. For the 30B model, IBM added a second fine-tuning phase that increased the share of agentic coding data while retaining about 16% replay data from the original mixture.

After fine-tuning, IBM applied a staged reinforcement-learning pipeline. All three models received foundational reinforcement learning for verifiable tasks and a final RLHF stage for preference and safety. The 8B and 30B models also received agentic reinforcement learning in three stages: software engineering, terminal operation and web search. IBM says those stages used real repositories, live shell environments and browsing tools, with rewards based on whether tasks were completed. The training used asynchronous GRPO, with separate generation and training workers, and relied on NeMo-RL and NeMo-Gym. The 3B model did not receive the agentic-RL block. The source also describes quantized releases in FP8, NVFP4, MXFP4 and multiple GGUF formats.

來源詳情: huggingface.co ↗

為什麼這很重要

The release provides unusually detailed visibility into how an openly licensed model family combines conventional pretraining with staged reinforcement learning for tool use. It also gives developers smaller models, quantized variants and OpenAI-compatible serving options that could make local or self-hosted reasoning and agentic workflows more practical. The reported benchmark results are IBM’s own evaluations, however, and the source does not establish independent replication, real-world reliability or broad availability beyond the described model releases.

The release matters because it makes the training process more inspectable than a typical model announcement. IBM provides stage-by-stage descriptions of the data, reward signals, rollout environments and optimization settings, including the distinction between verifiable rewards, judge-based rewards and agentic outcome rewards. That documentation is useful to researchers and developers evaluating whether tool use should be learned through ordinary instruction tuning, reinforcement learning or a combination. It also makes clear that the three model sizes are not simply scaled versions of one another: the 8B and 30B versions receive additional training intended to teach actions in environments, while the 3B version follows a shorter path.

The Apache 2.0 license and the release of quantized variants may broaden the practical options for organizations that want to run models under their own infrastructure. The article describes support for Transformers, vLLM and SGLang, an OpenAI-compatible endpoint, and integration instructions for OpenCode, Pi and OpenHands. These features could reduce adaptation work for teams already using compatible serving and agent frameworks. They do not, by themselves, show that the models can run economically on ordinary consumer hardware. The source identifies large-scale training on an NVIDIA GB200 NVL72 cluster hosted by CoreWeave and gives extensive distributed-training details, but it does not provide complete inference-cost comparisons or hardware requirements for each quantized model.

IBM’s reported results suggest capability increases with model size, especially on the listed reasoning, long-context and agentic coding evaluations. The source reports, for example, SWE-Bench Verified scores of 47.67 for 8B and 57.00 for 30B, RULER 128K scores of 71.41 and 81.38, and AIME25 scores of 86.67 and 89.17. These numbers are useful as a record of the company’s evaluation claims, but they are not independent evidence. The article does not describe a third-party audit, confidence intervals, contamination analysis, comparative testing against current alternatives or the operational failure modes encountered in the environments. It also does not establish how the models behave when tools return misleading information or when tasks have safety consequences.

接下來看什麼

The key questions are whether Granite 4.2’s reported capabilities hold up in independent testing, how much performance is lost in its FP8, FP4 and GGUF variants, and how reliably the 8B and 30B models act in less controlled environments. Users should also examine licensing and deployment details, hardware requirements, tool-call safety and the limits of the models’ long-context and agentic performance before treating benchmark scores as production evidence.

Independent evaluations should be the first follow-up. Reviewers will need to reproduce the listed benchmarks where possible, compare the models with similarly sized systems, inspect the evaluation prompts and determine whether the training data overlaps with test sets. Agentic results deserve especially careful scrutiny because success rates can depend heavily on the harness, repository selection, hidden tests, browser tools and task limits. IBM’s source reports the environments and some configuration details, but it does not provide enough information here to determine how representative they are of production work.

Deployment constraints are another important unknown. The source says Granite 4.2 can be served through OpenAI-compatible infrastructure and offers several quantization formats, but it does not state memory requirements, throughput, latency, energy use or quality degradation for each variant. The 512K context claim also needs practical testing: long context capacity is not the same as reliable retrieval or reasoning over every part of a very long input. Developers should measure performance on their own workloads and verify how thinking controls, history truncation and tool-call parsing behave in their chosen serving stack.

Safety and governance questions remain open as well. IBM says the final RLHF stage includes preference optimization, jailbreak resistance, appropriate refusals and a penalty for overly verbose reasoning. The article does not provide detailed safety results, refusal-error rates, privacy analysis, security testing or evidence about behavior outside the training environments. It also does not say whether chain-of-thought content is always exposed, filtered or handled differently across interfaces. Before using Granite 4.2 in consequential applications, organizations will need their own testing of tool permissions, data handling, prompt injection, auditability and human review. The immediate significance of the release is therefore both the models themselves and the more transparent account of the engineering choices behind them; the extent of their real-world advantage remains to be established.

相關指引和測驗