Nanbeige4.3-33B:面向编码与办公代理的开源权重模型(512K)
Nanbeige 发布 Apache-2.0 的 Nanbeige4.3-33B 开源权重模型,定位编码与办公风格代理,上下文 512K;实验室自测表在 Terminal-Bench 2.1、DeepSWE 等多项上高于 Qwen3.8-27B,BF16 约 66GB。
本地跑编码代理时,权重许可与上下文长度往往比排行榜分数更先卡人——能商用、能拉长上下文,才谈得上把文档与仓库塞进同一会话。Nanbeige 放出 Nanbeige4.3-33B:开源权重、定位编码与办公风格代理,上下文 512K,许可 Apache 2.0,面向希望自托管推理栈的团队。
据 @A_Radwan(约 2026-10-29)转述实验室自测表(各模型取其最高 thinking effort):相对 Qwen3.8-27B,它在 Terminal-Bench 2.1(80.9 vs 73.0)、DeepSWE(65.5 vs 42.5)、Toolathlon-Verified(70.4 vs 68.5)等项领先;在 SWE-Bench Pro(60.3 vs 61.7)与 GDPval(84.8 vs 86.4)等项落后;帖文称 20 行里有 17 行对其有利。BF16 体积约 66GB,本地更现实的路径是量化构建;尚无官方 Ollama 库页面,社区有 Hugging Face 上的 GGUF。
这是代理向小模型选型参考,而非通用「最强开源」结论,更不宜直接外推到通用聊天或闭源旗舰对比。原帖基准为实验室自报,且未给出第三方复现协议;上线前应在目标工具调用集与办公文档任务上自测,并确认量化精度对长上下文的影响,避免只看单项榜首做采购决策或替换已稳定的生产路由。仓库与社区量化包是否跟进,也需要单独核对。
Nanbeige4.3-33B Open-Weights for Coding and Office Agents
Nanbeige released Nanbeige4.3-33B: an open-weights 33B model for coding and office-style agents, with a 512K context window.
By the lab's own table (each model at its highest thinking effort), it beats Qwen3.8-27B on:
- Terminal-Bench 2.1: 80.9 vs 73.0
- DeepSWE: 65.5 vs 42.5
- Toolathlon-Verified: 70.4 vs 68.5
It trails on:
- SWE-Bench Pro: 60.3 vs 61.7 (Qwen3.8-27B)
- GDPval: 84.8 vs 86.4 (Qwen3.8-27B)
- GPQA Diamond: 89.5 vs 91.3 (Gemma4-31B)
17 of the 20 rows go its way. The weights are 66 GB at BF16, so a quantized build is the realistic local route; there is no Ollama library page, only community GGUFs on Hugging Face.
Licence: Apache 2.0.
For a coding agent on your own hardware, would you take the 33B that wins most rows or the 27B that still wins on SWE-Bench Pro?