[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-nanbeige-4-3-33b-open-agent-teacher-model":3,"topics-all":38,"news-related-9597f154-5515-457e-9590-8b32347ea750":57},{"id":4,"title":5,"summary":6,"content":7,"original_url":8,"source_id":9,"tags":10,"translations":24,"news_slug":31,"published_at":32,"created_at":33,"modified_at":34,"is_published":35,"publish_type":36,"image_url":14,"view_count":37},"9597f154-5515-457e-9590-8b32347ea750","Nanbeige4.3-33B:33B教师模型自测80.9分","Nanbeige 发布 33B 稠密开源模型 Nanbeige4.3-33B,Apache 2.0 许可、512K 上下文,定位 4.3 系列旗舰教师模型。厂商自测 Terminal-Bench 2.1 得 80.9 分,领先 Qwen3.8-27B,但 SWE-Bench Pro 等三项落后,暂无第三方复现。","开放权重这一轮的新玩法，是先把「老师」放出来。最近一个例子来自 Nanbeige 实验室：他们没有直接扔出小模型冲榜，而是先发了一个 33B 的稠密模型当 4.3 系列的教师模型，打算之后再把能力蒸馏进一个 3B 版本。这跟行业里常见的「先发小模型冲性价比」顺序正好相反。\n\n先看这个教师模型是什么。[模型卡](https:\u002F\u002Fhuggingface.co\u002FNanbeige\u002FNanbeige4.3-33B)显示，Nanbeige4.3-33B 是 Apache 2.0 许可的稠密模型，Hugging Face 元数据显示约 331 亿参数，BF16 权重约 66GB，上下文做到 524,288 tokens（512K），中英文双语。指令版和 Base 版权重同期放出，HF 页面上已经出现 44 个社区量化版本和 5 个二次微调版本，架构标签是 Llama。Nanbeige 之前做的是 4.1-3B、4.2-3B 这类小模型，这次 33B 的定位写得很直白：它是「旗舰教师」，小模型蒸馏管线在它的下游。\n\n## 榜单：赢在长程，输在三处\n\n厂商的对比组是 Qwen3.8-27B、Muse-Glimmer-30B 和 Gemma4-31B，全部拉到最高思考强度，同一套推理超参与 harness。几个关键行：\n\n- **Terminal-Bench 2.1**（Terminus 2 harness）：80.9，对手 73.0 \u002F 61.8 \u002F 47.2\n- **Terminal-Bench 4.0**（mini-swe-agent v2.4.6）：23.8，对手 6.3 \u002F 1.6 \u002F 0.0\n- **DeepSWE**（mini-swe-agent v2.4.6）：65.5，对手 42.5 \u002F 4.4 \u002F 2.7\n- **SWE-Bench Pro**：60.3，反而低于 Qwen3.8-27B 的 61.7\n- **GDPval**：84.8，低于 86.4\n- **GPQA Diamond**：89.5，低于 Gemma4-31B 的 91.3\n\n差距的分布比单个数字更有意思：长程 agent 任务上被拉开十几到二十几分，Terminal-Bench 4.0 甚至是 23.8 对 6.3；而在纯推理侧，AIME 2026 是 95.2 对 94.5，GPQA Diamond 89.5 对 89.2，几乎贴脸。厂商把输的三项都摆进了表里，这比大多数发布会表格诚实。\n\n## Cowork 那半张表可能更值钱\n\n比编程榜单更值得看的，是被标成 Agentic Cowork 的那一半：Automation Bench 54.2 对 38.3，JobBench 44.8 对 34.5（模型卡说 35 个职业里多数领先），YC-bench 上按美元计价是 1,025,437 对 641,290，Toolathlon-Verified 70.4 对 68.5，Claw-Eval 78.3（只算 157 道单轮任务），Tau3-Bench 48.8 对 48.5，DeepPlanning 63.3 对 49.2，EnterpriseOps-Gym 41.0 对 40.2。\n\n这半张表测的是「像办公助理一样跨工具干活」，而不是刷代码题。但它也更脆：这些基准高度依赖 harness、暴露了哪些工具、用谁来当评委——SWE-Atlas-QnA 一项就是由 GPT-6 Sol 按 1,304 条任务级 rubric 判分（50.8 对 41.9）。而且对手是厂商自己挑的，模型卡没跟更大的开源 MoE 比，更没跟闭源前沿比。\n\n别忘了基座版：Nanbeige4.3-33B-Base 在八项基座基准上全面压过 Gemma4-31B-Base，MMLU-Pro 66.9 对 59.3，GPQA 62.0 对 41.5。\n\n## 跑起来：66GB 权重和 512K 上下文的现实\n\nBF16 权重约 66GB，单张 80GB 卡塞进 512K 上下文的 KV cache 后相当紧张，官方示例是四卡张量并行，SGLang 和 vLLM 都直接给了命令（`--tp-size 4`、`--reasoning-parser qwen3`、`--tool-call-parser qwen3_coder`）。chat template 有两个关键开关：`enable_thinking` 默认打开；`preserve_thinking` 官方建议在多轮工具调用、办公任务和 code agent 场景保留，普通问答关掉；工具调用格式推荐 `xml`，`json` 作为兼容选项。参数建议分两档：agent 与工具类任务用 temperature 1.0、top_p 0.95、max_new_tokens 65,536；推理与闲聊用 temperature 0.6、top_p 0.95、max_new_tokens 262,144。另外 Transformers 示例带 `trust_remote_code=True`，在放密钥的机器上跑之前值得先扫一遍仓库代码。\n\n## 所以呢\n\n先把话说清楚：这张表里的每一个数字都是厂商自测，目前没有第三方复现，几个转述站点也都把它标注为厂商自报结果。评测口径对 agent 榜单的影响尤其大——换 harness 数字就会动。\n\n那它值不值得试？我的建议是别拿公开榜单当结论，挑自己两三个重复出现的任务（工单分拣、从表格起草报告），在你自己用的 harness 里跑一遍，比的是「每个完成任务的成本」而不是分数。分数高五分但重试次数翻倍，未必更便宜。\n\n更根本的问题是：这条「先立教师、再压学生」的路线，成败其实不在这个 33B 有多强，而在那个还没发布的 Nanbeige4.3-3B 能不能把老师的东西学个七七八八。33B 是手段，3B 才是目的。\n","https:\u002F\u002Fhuggingface.co\u002FNanbeige\u002FNanbeige4.3-33B","24d5c6c5-6573-4180-a1fd-f1459842d1af",[11,15,18,21],{"id":12,"name":13,"slug":13,"description":14,"color":14},"6ad31a14-c0da-42df-81fd-564281f768db","agentic-ai",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"120fa59a-ff6f-4537-9bf5-f818df636a0e","benchmark",{"id":19,"name":20,"slug":20,"description":14,"color":14},"7e89b5cc-57db-4f37-bc6d-28919a73931c","model-release",{"id":22,"name":23,"slug":23,"description":14,"color":14},"b9bd9039-fcdb-41a8-b85b-fc1587def2b9","open-source",[25],{"id":26,"lang":27,"title":28,"summary":29,"content":30},"093f3e21-5735-41f7-bfcd-56aede8a16ec","en","Nanbeige4.3-33B: open teacher model, 80.9 Terminal-Bench","Nanbeige's 33B open model (Apache 2.0, 512K context) is the teacher for its 4.3 series. Self-test: 80.9 Terminal-Bench 2.1, ahead of Qwen3.8-27B.","The new pattern in open-weight releases is to ship the teacher first. Nanbeige did not push out a small model to chase leaderboard spots; it released a 33B dense model as the flagship teacher of the 4.3 series, with a 3B student planned to be distilled from it later. That inverts the usual release order, where small models went out first to win on cost-effectiveness.\n\nHere is what the teacher actually is. The [model card](https:\u002F\u002Fhuggingface.co\u002FNanbeige\u002FNanbeige4.3-33B) describes Nanbeige4.3-33B as a dense model under the Apache 2.0 license, with Hugging Face metadata showing roughly 33.1 billion parameters, BF16 weights around 66GB, a context window of 524,288 tokens (512K), and support for English and Chinese. Both the instruct checkpoint and a Base checkpoint were released at the same time, and the Hugging Face page already carries 44 community quantizations and 5 third-party finetunes, with a Llama architecture tag. Nanbeige previously built small models such as 4.1-3B and 4.2-3B. This time the 33B is stated plainly as the flagship teacher, with the small-model distillation pipeline downstream of it.\n\n## The leaderboard: big wins on long-horizon work, three losses\n\nThe vendor's comparison set is Qwen3.8-27B, Muse-Glimmer-30B and Gemma4-31B, all run at their highest thinking effort with identical inference hyperparameters and harness per benchmark. The key rows:\n\n- **Terminal-Bench 2.1** (Terminus 2 harness): 80.9, against 73.0 \u002F 61.8 \u002F 47.2\n- **Terminal-Bench 4.0** (mini-swe-agent v2.4.6): 23.8, against 6.3 \u002F 1.6 \u002F 0.0\n- **DeepSWE** (mini-swe-agent v2.4.6): 65.5, against 42.5 \u002F 4.4 \u002F 2.7\n- **SWE-Bench Pro**: 60.3, below Qwen3.8-27B's 61.7\n- **GDPval**: 84.8, below 86.4\n- **GPQA Diamond**: 89.5, below Gemma4-31B's 91.3\n\nThe distribution of those gaps is more interesting than any single number. On long-horizon agent tasks the margin runs into the teens or twenties, with Terminal-Bench 4.0 landing at 23.8 versus 6.3. On pure reasoning the gap nearly vanishes: AIME 2026 is 95.2 against 94.5, and GPQA Diamond is 89.5 against 89.2. The vendor put all three losses in the table, which is more candid than most launch-day benchmark sheets.\n\n## The Agentic Cowork half may matter more\n\nMore interesting than the coding leaderboards is the half of the table labeled Agentic Cowork: Automation Bench 54.2 against 38.3, JobBench 44.8 against 34.5 (the model card says it beats the baseline in most of the 35 occupations evaluated), YC-bench priced in dollars at 1,025,437 against 641,290, Toolathlon-Verified 70.4 against 68.5, Claw-Eval 78.3 (counting only the 157 single-turn tasks in the General subset), Tau3-Bench 48.8 against 48.5, DeepPlanning 63.3 against 49.2, and EnterpriseOps-Gym 41.0 against 40.2.\n\nThat half tests whether a model can do multi-step work across business tools like an office assistant, rather than solving coding puzzles. It is also the more fragile half. These suites depend heavily on the harness, on which tools get exposed, and on who grades: SWE-Atlas-QnA alone is graded by a rubric judge, GPT-6 Sol, against 1,304 task-specific rubrics (50.8 against 41.9). And the rivals were picked by the lab itself. The card does not compare against larger open MoE models, and it does not compare against closed frontier models either.\n\nThe base checkpoint deserves a mention too: Nanbeige4.3-33B-Base beats Gemma4-31B-Base across all eight base-model benchmarks, with MMLU-Pro 66.9 against 59.3 and GPQA 62.0 against 41.5.\n\n## Running it: 66GB of weights and a 512K context\n\nAt BF16 the weights are about 66GB, which makes a single 80GB GPU tight once you add KV cache for a 512K context, and the official example uses four-way tensor parallelism. Both SGLang and vLLM get explicit commands (`--tp-size 4`, `--reasoning-parser qwen3`, `--tool-call-parser qwen3_coder`). Two chat-template switches matter: `enable_thinking` is on by default, and `preserve_thinking` is recommended for multi-turn tool use, office tasks and code-agent workflows but turned off for ordinary chat and question answering. The recommended tool-call format is `xml`, with `json` kept for compatibility. Inference settings split into two profiles: for agentic and tool-use tasks, temperature 1.0, top_p 0.95 and max_new_tokens 65,536; for reasoning and chat, temperature 0.6, top_p 0.95 and max_new_tokens 262144. The Transformers example also loads with `trust_remote_code=True`, so it is worth reading the repository code before running it on a machine holding secrets.\n\n## So what\n\nLet us be clear about one thing first: every number in that table is the lab's own test run. There is no third-party reproduction at the time of writing, and the outlets covering it all label the results as vendor-reported. Harness choice matters enormously for agent leaderboards, and the numbers move when you change it.\n\nIs it worth trying? My advice is not to treat a public leaderboard as the conclusion. Pick two or three recurring tasks from your own work (triaging a ticket queue, drafting a report from a spreadsheet), run them through the harness you actually use, and compare the cost per completed task rather than the score. A model that scores five points higher but needs twice the retries is not cheaper.\n\nThe deeper question is this: the success of the \"build a teacher first, then compress a student\" route does not really depend on how strong this 33B is. It depends on whether the not-yet-released Nanbeige4.3-3B can absorb what the teacher taught it. The 33B is the means; the 3B is the point.\n","nanbeige-4-3-33b-open-agent-teacher-model","2026-10-26T01:15:00Z","2026-10-26T01:23:08.920263Z","2026-10-26T01:23:08.920273Z",true,"agent",90,[39,48],{"slug":40,"tag_slug":40,"title_zh":41,"title_en":42,"intro_zh":43,"intro_en":44,"id":45,"is_active":35,"created_at":46,"modified_at":47},"ai-for-science","AI for Science 2026：从 UniPert 到 GPT-Rosalind 的硬核进化","AI for Science 2026: from UniPert to GPT-Rosalind","生命科学、化学材料、物理世界模型——AI 正在从\"语言工具\"变成\"实验伙伴\"。本专题收录 AI 在三大科学方向的关键节点：UniPert 统一基因与化学扰动空间、GPT-Rosalind 端到端生命科学推理、达摩院 AI 智能体 28 小时找到 4 种超导新材料、Anthropic Claude Science 把工作台做成标准品。","From language tool to lab partner — AI is reshaping life sciences, chemistry\u002Fmaterials, and physical world models. This topic covers the key milestones: UniPert unifying genetic-chemical perturbation spaces, GPT-Rosalind's end-to-end life-sciences reasoning, DAMO's AI agent discovering 4 superconducting materials in 28 hours, and Anthropic's Claude Science workbench going mainstream.","988a4300-5fab-41c4-b5d8-63711a2dc757","2026-09-10T01:34:15.296649Z","2026-09-10T01:34:15.296663Z",{"slug":49,"tag_slug":49,"title_zh":50,"title_en":51,"intro_zh":52,"intro_en":53,"id":54,"is_active":35,"created_at":55,"modified_at":56},"h3-series","MiniMax H3 系列：从开源权重到 35 倍吞吐","MiniMax H3 Series: from open weights to 35x throughput","MiniMax H3 自 2026 年 8 月开源以来节奏密集：官方把生成、参考与编辑收回一个模型；ComfyUI 当天压进 RTX 3060；摩尔线程 3 小时完成国产 GPU 适配；fal 后训练版把吞吐拉到 35 倍；FastH3 蒸馏再砍推理成本。本专题持续追踪 H3 的发布—开源—蒸馏—部署全链路。","Since MiniMax open-sourced H3 in August 2026 the pace has been relentless: one unified omni-modal model, same-day ComfyUI support down to an RTX 3060, a 3-hour Day-0 port to Moore Threads GPUs, fal's post-trained H3 Max at 35x throughput, and FastH3 distillation cutting inference cost further. This topic tracks the full H3 chain — release, open weights, distillation, deployment.","83ef0daa-3c31-4cb3-86ed-e5ee58654d5f","2026-09-08T07:33:19.942193Z","2026-09-08T07:33:19.942209Z",{"items":58},[59,64,69,74,79,84],{"id":60,"title":61,"news_slug":62,"published_at":63},"983fb4d1-6c63-4828-adbf-a59c57e02b64","Perplexity开源决策模型:总分微胜Jev","perplexity-decider-v1-27b-open-source","2026-10-02T21:20:00+00:00",{"id":65,"title":66,"news_slug":67,"published_at":68},"c94766df-827e-4e4e-a006-b6639ec76722","DeepSeek V4-Flash-0731 转正观察:权重不动,后训练把 Agent 分数打到 V4-Pro 之上","deepseek-v4-flash-0731-agent-benchmark-official-aug2026","2026-08-01T02:00:00+00:00",{"id":70,"title":71,"news_slug":72,"published_at":73},"dcc6132a-e432-4ef3-9555-5412d692cd90","RSIHub 开源:先冻结评测器,再谈自我进化","rsihub-agent-evolution-frozen-evaluator","2026-10-26T07:20:28+00:00",{"id":75,"title":76,"news_slug":77,"published_at":78},"7d735b5c-c7a8-4a8d-afac-abd3034170e7","Sage-A1.5开源:35B智能体GAIA 95.1","sage-a1-5-agentic-scientific-discovery","2026-10-24T23:10:02+00:00",{"id":80,"title":81,"news_slug":82,"published_at":83},"a744ef64-ee58-41ea-892f-1c2e5ba50bb1","微信视觉团队开源WeVisOne-VL:检测到抠图全包","wevisone-vl-unified-visual-perception","2026-10-23T17:07:19+00:00",{"id":85,"title":86,"news_slug":87,"published_at":88},"aa718f3c-14cf-4984-8762-794eaaee8c16","MAI-Gym:9B 智能体在 2 万个并行环境里做在线 RL","mai-gym-gui-agent-online-rl-9b","2026-10-21T09:30:00+00:00"]