第一项始终是图钉的来源。总体可信度是各份资料对上方所用开始和结束时间支持程度的加权平均;资料每比最新的一份旧 180 天,权重减半。
IBM在Hugging Face[5]上的公告《Granite 4.2 LLMs: How They're Built》以及ibm-granite/granite-4.2-8b模型卡都将发布日期标为2026年8月25日;OpenRouter[2]于8月31日上架8B模型,即该路由器添加它的日期。
Router listing of the 8B: 131,072-token context and $0.06 per 1 million input tokens and $0.25 per 1 million output tokens.[1]
Independent trade report: Granite 4.1 (April) stressed answering without long reasoning chains, 4.2 makes reasoning optional; only the 8B and 30B got the agentic RL stage; 1 trillion synthetic code tokens from CodeAlchemy; IBM's 30B resolves 57% on SWE-bench Verified.[1]
Releases are dated 25 August 2026; flags that the benchmark figures are IBM's own runs, that the models are text-only, and that Qwen3.8-27B leads on published coding benchmarks.[1]
IBM's model card for the 8B: dense decoder-only, post-trained from Granite-4.1-8B-Base, Apache 2.0, 12 tested languages, thinking / non-thinking / low-effort modes, GRPO then RLHF post-training on a GB200 NVL72 cluster at CoreWeave, and the 3B/8B/30B results table.[1]
有遗漏或错误吗?用你自己的话说明:能佐证此图钉的链接、不同的开始或结束日期及理由,或缺失、有误的信息。AI 会对照此图钉的来源进行核实,搜索更好的来源,并添加任何支持你说法的页面。图钉自身的来源仍然最重要。AI 也会查看图片:显示的是别的东西或显示效果差的图片会被移到后面或替换。
原标题: Granite 4.2 8B, 3B and 30B Released
| 基准 | Granite 4.2 3B | Granite 4.2 8B | Granite 4.2 30B |
|---|---|---|---|
| SWE-bench Verified(智能体编码) | 不适用 | 47.67 | 57.00 |
| SWE-bench Multilingual(智能体编码) | 不适用 | 30.78 | 41.89 |
| SWE-bench Pro(智能体编码) | 不适用 | 19.11 | 33.29 |
| Terminal-Bench 2.1(智能体编码) | 不适用 | 20.56 | 29.24 |
| τ³-bench平均分(智能体,通用) | 45.78 | 58.06 | 62.00 |
| AIME25 (推理) | 78.33 | 86.67 | 89.17 |
| HMMT Feb25 (推理) | 66.67 | 78.33 | 89.17 |
| GPQA (推理) | 54.80 | 64.14 | 66.41 |
| LiveCodeBench v6 (推理) | 69.71 | 73.24 | 75.77 |
| SciCode (推理) | 24.11 | 36.09 | 38.76 |
| MMLU-Pro (对话与指令遵循) | 67.84 | 74.04 | 77.60 |
| Arena-Hard-V2 (对话与指令遵循) | 34.96 | 65.19 | 67.93 |
| IFBench, prompt (对话与指令遵循) | 74.33 | 79.33 | 77.17 |
| RULER 64K (长上下文) | 67.52 | 80.99 | 89.96 |
| RULER 128K (长上下文) | 55.30 | 71.41 | 81.38 |