[{"data":1,"prerenderedAt":59},["Reactive",2],{"news-story-news-20261025-nanbeige-4-3-33b-open-agentic-model":3},{"id":4,"title_zh":5,"title_en":6,"summary_zh":7,"summary_en":8,"takeaways_zh":9,"takeaways_en":15,"content_zh":21,"content_en":22,"external_links":23,"source":31,"category":38,"impact_score":39,"views_count":40,"likes_count":41,"related_entities":42,"tags":52,"image_url":58},"news-20261025-nanbeige-4-3-33b-open-agentic-model","BOSS 直聘南北阁实验室开源 Nanbeige4.3-33B：33B 稠密模型、512K 上下文、Apache 2.0，官方称 Terminal-Bench 2.1 达 80.9 超过 Qwen3.8-27B","BOSS Zhipin's Nanbeige lab open-sources Nanbeige4.3-33B, a 33B dense agent model with 512K context under Apache 2.0, reporting 80.9 on Terminal-Bench 2.1 vs 73.0 for Qwen3.8-27B","BOSS 直聘（看准科技）旗下南北阁大模型实验室 @nanbeige 在 UTC 2026-10-25 12:11 发布 Nanbeige4.3-33B（[原帖](https://x.com/nanbeige/status/2114328933181333798)），模型和 Base 版已在 Hugging Face 以 Apache 2.0 开放权重（[模型卡](https://huggingface.co/Nanbeige/Nanbeige4.3-33B)）。这是一个 Llama 架构的 33B 稠密模型，支持 512K 上下文，定位为 Nanbeige4.3 系列的「教师模型」，后续会蒸馏出 Nanbeige4.3-3B 等小模型。官方模型卡在同一套 harness、最高思考强度下对比 Qwen3.8-27B、Muse-Glimmer-30B、Gemma4-31B：Terminal-Bench 2.1 为 80.9（Qwen3.8-27B 73.0），DeepSWE 65.5（42.5），但 SWE-Bench Pro 60.3 略低于 Qwen 的 61.7，GDPval 84.8 也低于 86.4。以上均为官方自测成绩，尚无第三方复现。","At 12:11 UTC on 2026-10-25, the Nanbeige LLM Lab of BOSS Zhipin (Kanzhun Limited) announced Nanbeige4.3-33B from its account @nanbeige ([post](https://x.com/nanbeige/status/2114328933181333798)), with the model and its Base version released as open weights on Hugging Face under Apache 2.0 ([model card](https://huggingface.co/Nanbeige/Nanbeige4.3-33B)). It is a 33B dense model on a Llama-style architecture with a 512K context window, positioned as the teacher model of the Nanbeige4.3 series that will be distilled into smaller models such as Nanbeige4.3-3B. The official model card compares it with Qwen3.8-27B, Muse-Glimmer-30B and Gemma4-31B using the same harness at highest thinking effort: 80.9 on Terminal-Bench 2.1 (Qwen3.8-27B 73.0) and 65.5 on DeepSWE (42.5), but 60.3 on SWE-Bench Pro, slightly below Qwen's 61.7, and 84.8 on GDPval vs 86.4. All figures are self-reported and not yet independently reproduced.",[10,11,12,13,14],"官方来源：[@nanbeige X 原帖](https://x.com/nanbeige/status/2114328933181333798) 与 [Hugging Face 模型卡](https://huggingface.co/Nanbeige/Nanbeige4.3-33B)，抓取时原帖约 616 赞、79 转发、3.8 万浏览","规格：33B 稠密、Llama 架构、512K 上下文、Apache 2.0，同时开放 Base 版","官方自测编码：Terminal-Bench 2.1 80.9、Terminal-Bench 4.0 23.8、DeepSWE 65.5、SWE-Bench Pro 60.3（Qwen3.8-27B 分别为 73.0、6.3、42.5、61.7）","官方自测协作类：Automation Bench 54.2、JobBench 44.8、GDPval 84.8（Qwen3.8-27B 为 38.3、34.5、86.4）","部署：支持 Transformers、SGLang、vLLM，官方示例用 4 卡张量并行；后续将蒸馏 Nanbeige4.3-3B",[16,17,18,19,20],"Official sources: [@nanbeige post on X](https://x.com/nanbeige/status/2114328933181333798) and the [Hugging Face model card](https://huggingface.co/Nanbeige/Nanbeige4.3-33B); the post had about 616 likes, 79 reposts and 38K views when we pulled it","Specs: 33B dense, Llama-style architecture, 512K context, Apache 2.0, with a Base checkpoint released too","Self-reported coding scores: Terminal-Bench 2.1 80.9, Terminal-Bench 4.0 23.8, DeepSWE 65.5, SWE-Bench Pro 60.3 (Qwen3.8-27B: 73.0, 6.3, 42.5, 61.7)","Self-reported cowork scores: Automation Bench 54.2, JobBench 44.8, GDPval 84.8 (Qwen3.8-27B: 38.3, 34.5, 86.4)","Deployment: Transformers, SGLang and vLLM, with official examples using 4-way tensor parallelism; Nanbeige4.3-3B distillation is next","### 核心背景与行业痛点\n30B 左右的开源模型是很多团队本地部署和私有化 Agent 的主力尺寸，但这个尺寸在终端操作、长链路编码和办公自动化这类 Agent 任务上一直明显落后于闭源大模型。BOSS 直聘（看准科技）旗下的南北阁大模型实验室此前以 Nanbeige4.2-3B 等小模型为主，这次在 UTC 10-25 12:11 通过官方账号发布 Nanbeige4.3-33B（[原帖](https://x.com/nanbeige/status/2114328933181333798)），思路是先做一个强的「教师模型」，再把 Agent 能力蒸馏到更小的模型里。\n\n### 架构亮点与底层机制\n根据 [Hugging Face 模型卡](https://huggingface.co/Nanbeige/Nanbeige4.3-33B)，Nanbeige4.3-33B 是 Llama 架构的 33B 稠密模型（不是 MoE），支持最长 524,288 token（512K）上下文，以 Apache 2.0 许可开放，并同时发布了 [Base 版](https://huggingface.co/Nanbeige/Nanbeige4.3-33B-Base)。聊天模板提供 `enable_thinking`（是否输出推理）和 `preserve_thinking`（多轮中是否保留之前的推理）两个开关，官方建议多轮工具调用、办公任务和代码 Agent 场景打开 `preserve_thinking`，工具调用格式推荐 XML。官方明确它是 Nanbeige4.3 系列的教师模型，下一步是 Nanbeige4.3-3B。\n\n### 权威 Benchmark 与实测跑分对比\n以下全部来自官方模型卡，在同一 harness、最高思考强度下与 Qwen3.8-27B 对比：Terminal-Bench 2.1（Terminus 2）80.9 对 73.0；Terminal-Bench 4.0 23.8 对 6.3；DeepSWE 65.5 对 42.5；SWE-Bench Pro 60.3 对 61.7（落后）；NL2Repo（用 Claude Code 评测）53.0 对 48.3；Automation Bench 54.2 对 38.3；JobBench 44.8 对 34.5；GDPval 84.8 对 86.4（落后）；GPQA Diamond 89.5 对 89.2，但低于 Gemma4-31B 的 91.3；AIME 2026 95.2 对 94.5。这些是厂商自测，尚无第三方复现；X 上已有个别用户反馈在 3D 生成提示词上首轮效果不理想，实际表现还要等更多社区实测。\n\n### 开发者实战落地与开箱指南\n权重可直接从 Hugging Face 下载，官方给出 Transformers、SGLang、vLLM 三种运行方式，SGLang 和 vLLM 示例都用 4 卡张量并行，并复用 `qwen3` 推理解析器和 `qwen3_coder` 工具调用解析器。做 Agent 场景时，官方建议温度 1.0、top_p 0.95、最大生成 65,536 token；推理和聊天建议温度 0.6。想在本地或私有环境跑编码 Agent 的团队，可以拿它和 Qwen3.8-27B 在自己的任务集上做对比，再决定是否替换。","### Background and the problem\nOpen models around 30B parameters are the workhorse size for local and private agent deployments, but they have lagged well behind closed frontier models on agent work such as terminal operations, long-horizon coding and office automation. The Nanbeige LLM Lab of BOSS Zhipin (Kanzhun Limited), previously known for small models like Nanbeige4.2-3B, announced Nanbeige4.3-33B from its official account at 12:11 UTC on 10-25 ([post](https://x.com/nanbeige/status/2114328933181333798)). The strategy is to build a strong teacher model first, then distill its agent abilities into smaller models.\n\n### How it works\nAccording to the [Hugging Face model card](https://huggingface.co/Nanbeige/Nanbeige4.3-33B), Nanbeige4.3-33B is a 33B dense model on a Llama-style architecture (not MoE) with up to 524,288 tokens (512K) of context, released under Apache 2.0 alongside a [Base checkpoint](https://huggingface.co/Nanbeige/Nanbeige4.3-33B-Base). The chat template has two switches: `enable_thinking` (whether to produce reasoning) and `preserve_thinking` (whether to keep earlier reasoning in multi-turn conversations). The lab recommends turning on `preserve_thinking` for multi-turn tool use, office tasks and code agents, and recommends the XML tool-call format. It is explicitly the teacher model for the Nanbeige4.3 series, with Nanbeige4.3-3B next.\n\n### Benchmarks and measured results\nAll of the following come from the official model card, compared with Qwen3.8-27B using the same harness at highest thinking effort: Terminal-Bench 2.1 (Terminus 2) 80.9 vs 73.0; Terminal-Bench 4.0 23.8 vs 6.3; DeepSWE 65.5 vs 42.5; SWE-Bench Pro 60.3 vs 61.7 (behind); NL2Repo (evaluated with Claude Code) 53.0 vs 48.3; Automation Bench 54.2 vs 38.3; JobBench 44.8 vs 34.5; GDPval 84.8 vs 86.4 (behind); GPQA Diamond 89.5 vs 89.2, though below Gemma4-31B at 91.3; AIME 2026 95.2 vs 94.5. These are vendor-reported and not yet independently reproduced, and at least one early user on X reported a weak first attempt on a 3D generation prompt, so real-world performance still needs more community testing.\n\n### What developers can do now\nWeights download directly from Hugging Face, and the lab documents Transformers, SGLang and vLLM setups; the SGLang and vLLM examples use 4-way tensor parallelism and reuse the `qwen3` reasoning parser and `qwen3_coder` tool-call parser. For agent workloads the lab suggests temperature 1.0, top_p 0.95 and up to 65,536 new tokens; for reasoning and chat, temperature 0.6. Teams running coding agents locally or in private environments can compare it against Qwen3.8-27B on their own task sets before switching.",[24,28],{"title":25,"url":26,"type":27},"Nanbeige4.3-33B model card on Hugging Face","https://huggingface.co/Nanbeige/Nanbeige4.3-33B","other",{"title":29,"url":30,"type":27},"Nanbeige4.3-33B-Base on Hugging Face","https://huggingface.co/Nanbeige/Nanbeige4.3-33B-Base",{"platform":32,"author_name":33,"author_handle":34,"author_avatar":35,"tweet_url":36,"published_at":37},"x","Nanbeige","@nanbeige","https://unavatar.io/x/nanbeige","https://x.com/nanbeige/status/2114328933181333798","2026-10-25T14:12:26.000Z","release",84,247,0,{"models":43,"agents":48,"vendors":49},[44,45,46,47],"Nanbeige4.3-33B","Qwen3.8-27B","Gemma4-31B","Muse-Glimmer-30B",[],[33,50,51],"BOSS Zhipin","Kanzhun",[44,53,54,55,56,57],"南北阁","开源模型","Agent","Terminal-Bench","BOSS直聘","",1793718872287]