围绕Querying 3这一话题，我们整理了近期最值得关注的几个重要方面，帮助您快速了解事态全貌。
首先，Sarvam 105B shows strong, balanced performance across core capabilities including mathematics, coding, knowledge, and instruction following. It achieves 98.6 on Math500, matching the top models in the comparison, and 71.7 on LiveCodeBench v6, outperforming most competitors on real-world coding tasks. On knowledge benchmarks, it scores 90.6 on MMLU and 81.7 on MMLU Pro, remaining competitive with frontier-class systems. With 84.8 on IF Eval, the model demonstrates a well-rounded capability profile across the major workloads expected of modern language models.
，推荐阅读WhatsApp网页版获取更多信息
其次，BenchmarkSarvam-30BGemma 27B ItMistral-3.2-24B-Instruct-2506OLMo 3.1 32B ThinkNemotron-3-Nano-30BQwen3-30B-Thinking-2507GLM 4.7 FlashGPT-OSS-20BGENERALMath50097.087.469.496.298.097.697.094.2Humaneval92.188.492.995.197.695.796.395.7MBPP92.781.878.358.791.994.391.895.3Live Code Bench v670.028.026.073.068.366.064.061.0MMLU85.181.280.586.484.088.486.985.3MMLU Pro80.068.169.172.078.380.973.675.0Arena Hard v249.050.143.142.067.772.158.162.9REASONINGGPQA Diamond66.5--57.573.073.475.271.5AIME 25 (w/ tools)80.0 (96.7)--78.1 (81.7)89.1 (99.2)85.091.691.7 (98.7)HMMT Feb 202573.3--51.785.071.485.076.7HMMT Nov 202574.2--58.375.073.381.768.3Beyond AIME58.3--48.564.061.060.046.0AGENTICBrowseComp35.5---23.82.942.828.3SWE-Bench Verified34.0---38.822.059.234.0Tau2 (avg.)45.7---49.047.779.548.7
来自产业链上下游的反馈一致表明，市场需求端正释放出强劲的增长信号，供给侧改革成效初显。
第三，To their credit, Lenovo seems to fully understand that distinction. They told us straight out: “10/10 isn’t the destination. From our perspective it’s the new baseline…. But the real opportunity is to go beyond the score. A perfect rating only matters if it leads to meaningful outcomes: quicker repairs, longer‑lasting devices, lower ownership costs, and less waste. Measuring success through customer experience and real‑world repair data will be just as important as external benchmarks. Ultimately, repairability will continue to evolve. As expectations, regulations, and technologies change, so must our approach.”
此外，I used to work at a vector database company. My entire job was helping people understand why they needed a database purpose-built for AI; embeddings, semantic search, the whole thing. So it's a little funny that I'm writing this. But here I am, watching everyone in the AI ecosystem suddenly rediscover the humble filesystem, and I think they might be onto something bigger than most people realize.
最后，Notably, one thing it does is taking the “right” directory as a template, from which it first un-applies the original diff, then applies the modified version the user edited. This is because the “left” directory is marked read-only by Jujutsu, and I didn’t want to mark files writable while being careful not to touch other attributes.
随着Querying 3领域的不断深化发展，我们有理由相信，未来将涌现出更多创新成果和发展机遇。感谢您的阅读，欢迎持续关注后续报道。