跳到主内容
@wquguru
精选88MarkTechPost(RSS)模型发布/更新

OpenBMB发布MiniCPM5-2B:端侧部署与工具调用能力突出

OpenBMB Releases MiniCPM5-2B: A 2.52B Dense Model Averaging 53.9 Across 34 Benchmarks and Built to Run On Device

原文
发到 X
推荐理由

端侧小模型在工具调用和Agent场景有明确突破,且完整开源了从数据到中间检查点的训练链路,适合做轻量化Agent部署的研究者和开发者参考。

OpenBMB has released MiniCPM5-2B, the second checkpoint in the MiniCPM5 series and the follow-up to MiniCPM5-1B. It is a dense causal language model with 2,516,756,480 parameters, of which 1,981,982,720 sit outside the embeddings. It uses 42 layers, grouped-query attention with 16 query heads and 2 key/value heads, and a native context window of 131,072 tokens. The architecture is standard LlamaForCausalLM, so mainstream engines load it with no custom kernels and no model-code fork.

OpenBMB 发布了 MiniCPM5-2B,这是 MiniCPM5 系列的第二个检查点,也是 MiniCPM5-1B 的后续版本。它是一个拥有 2,516,756,480 个参数的密集因果语言模型,其中 1,981,982,720 个参数位于嵌入层之外。该模型采用 42 层结构,使用具有 16 个查询头和 2 个键/值头的分组查询注意力机制,原生上下文窗口为 131,072 个 token。其架构为标准 LlamaForCausalLM,因此主流引擎无需自定义内核或模型代码分支即可加载它。

Is it deployable? Yes. The weights are Apache 2.0 and run through vLLM, SGLang, Transformers, llama.cpp, Ollama, LM Studio, MLX and FlagOS.

可以部署吗?可以。权重采用 Apache 2.0 许可证,并兼容 vLLM、SGLang、Transformers、llama.cpp、Ollama、LM Studio、MLX 和 FlagOS。

What the benchmark table actually shows

基准测试表实际展示了什么

OpenBMB compares MiniCPM5-2B against LFM2.5-2.6B, Qwen3.5-2B and Gemma-4-E2B-it in the same size class, and lists Qwen3.5-4B, granite-4.2-3B, Nemotron-3-Nano-4B, Gemma-4-E4B-it and LFM2.5-8B-A1B for reference. Across 34 benchmark rows it averages 53.9. The best baseline in that set is Qwen3.5-4B at 51.1, then granite-4.2-3B at 42.7 and LFM2.5-2.6B at 33.2.

OpenBMB 将 MiniCPM5-2B 与同量级的 LFM2.5-2.6B、Qwen3.5-2B 和 Gemma-4-E2B-it 进行对比,并列出了 Qwen3.5-4B、granite-4.2-3B、Nemotron-3-Nano-4B、Gemma-4-E4B-it 和 LFM2.5-8B-A1B 作为参考。在 34 项基准测试中,其平均分为 53.9。该组中表现最好的基线模型是 Qwen3.5-4B(51.1 分),其次是 granite-4.2-3B(42.7 分)和 LFM2.5-2.6B(33.2 分)。

On code reasoning MiniCPM5-2B posts 69.1 on LiveCodeBench v6 against 56.4, and 46.4 on SWE-bench Verified against 33.6. Tool use is the widest margin: 97.1 on τ²-Bench Telecom, 66.6 on BFCL v4, and 20.8 on τ³-Bench Banking against 6.8. Long context is split, with 68.1 on NoLiMa against 43.5, but 59.0 on AA-LCR against 61.0 and 43.7 on LongBench v2 against 47.3. General knowledge is where the size gap shows: 70.8 on MMLU-Pro against 78.0, and 8.9 on Humanity’s Last Exam against 9.9. OpenBMB marks rows sourced from Artificial Analysis separately from internally reproduced ones.

在代码推理方面,MiniCPM5-2B 在 LiveCodeBench v6 上得分为 69.1(对比值为 56.4),在 SWE-bench Verified 上得分为 46.4(对比值为 33.6)。工具使用方面的优势最为显著:在 τ²-Bench Telecom 上得分为 97.1,在 BFCL v4 上得分为 66.6,在 τ³-Bench Banking 上得分为 20.8(对比值为 6.8)。长上下文能力表现不一:在 NoLiMa 上得分为 68.1(对比值为 43.5),但在 AA-LCR 上得分为 59.0(低于对比值的 61.0),在 LongBench v2 上得分为 43.7(低于对比值的 47.3)。通用知识领域体现了参数量差距:在 MMLU-Pro 上得分为 70.8(低于对比值的 78.0),在 Humanity’s Last Exam 上得分为 8.9(低于对比值的 9.9)。OpenBMB 将源自 Artificial Analysis 的行与内部复现的行分开标记。

Training recipe: SFT, then RL, then on-policy distillation

训练配方:监督微调(SFT),然后强化学习(RL),然后在线策略蒸馏

Training follows the UltraData tiered data management method described in original research. Base training runs stable and decay phases, then mid-training adapts the model to the target data distribution. Post-training starts with 400B tokens of deep-thinking SFT, then trains specialised RL teachers for math, code, agentic tasks and writing using the critic-based JustRL II algorithm.

训练遵循原始研究中描述的 UltraData 分层数据管理方法。基础训练运行稳定衰减阶段,随后中期训练使模型适应目标数据分布。后训练阶段首先使用 4000 亿 token 的深度思考数据进行监督微调,然后使用基于批评家的 JustRL II 算法为数学、代码、智能体任务和写作训练专门的强化学习教师模型。

The final step is on-policy distillation. OPD merges 16 RL experts, five of them agentic, into a single shipped model. At each response position it computes full-vocabulary reverse KL divergence between student and teacher logits as the advantage estimate, replacing the verification-based advantage. It reuses the RL prompts as distillation data, so no new corpus is built. OpenBMB measures the RL plus OPD stage at 10.96 average points on reasoning and general benchmarks and 6.96 points on agentic ones.

最后一步是策略内蒸馏(On-Policy Distillation,OPD)。OPD 将 16 个强化学习专家合并为一个发布模型,其中五个为智能体(agentic)专家。在每个响应位置,它计算学生模型与教师模型 logits 之间的全词表反向 KL 散度,作为优势估计值,从而取代基于验证的优势估计。它复用强化学习提示词作为蒸馏数据,因此无需构建新的语料库。OpenBMB 在推理和通用基准测试上测得 RL 加 OPD 阶段的平均分为 10.96 分,在智能体任务上为 6.96 分。

The data is open too

数据也是开源的。

Alongside the weights, OpenBMB released Ultra-FineWeb, Ultra-FineWeb-L3, UltraX, UltraData-Code, UltraData-Math, UltraData-SFT-2605, UltraData-SFT-Agent-2609 with 500K agent samples, and UltraData-RL-2609 with more than 80K RL samples. Intermediate checkpoints are published as well, covering Base, Midtrain and SFT-only, so the contribution of each stage can be measured directly.

除了权重之外,OpenBMB 还发布了 Ultra-FineWeb、Ultra-FineWeb-L3、UltraX、UltraData-Code、UltraData-Math、包含 50 万智能体样本的 UltraData-SFT-2605、包含超过 8 万个强化学习样本的 UltraData-RL-2609,以及 UltraData-SFT-Agent-2609。同时发布了中间检查点,涵盖 Base、Midtrain 和仅 SFT 阶段,以便直接衡量每个阶段的贡献。

Summary

总结

MiniCPM5-2B is a credible on-device option for agentic and tool-calling workloads, not a general knowledge model. Its advantage is clearest on tool use, coding agents and NoLiMa-style long-context retrieval, and it trails larger models on MMLU-Pro, GPQA-Diamond and MATH-500. The open data and intermediate checkpoints make the RL plus OPD claim checkable, which matters more than the headline average.

MiniCPM5-2B 是适用于智能体和工具调用负载的设备端可信选项,而非通用知识模型。其在工具使用、编码智能体和 NoLiMa 风格长上下文检索方面的优势最为明显,但在 MMLU-Pro、GPQA-Diamond 和 MATH-500 上落后于更大的模型。开源数据和中间检查点使得对“RL 加 OPD”这一主张的可验证性成为可能,这比 headline average(总体平均分)更为重要。

Key Takeaways

关键要点

  • 2.52B dense model, 131,072 token context, Apache 2.0, standard Llama architecture.
  • Averages 53.9 across 34 benchmarks, ahead of Qwen3.5-4B at 51.1.
  • Strongest on tool use, coding agents and long-context retrieval; weakest on knowledge.
  • Post-training pairs 400B SFT tokens with RL teachers and on-policy distillation.
  • Pre-training, SFT and RL datasets ship alongside the weights.
  • 25.2 亿参数的稠密模型,131,072 token 上下文窗口,Apache 2.0 许可证,标准 Llama 架构。
  • 在 34 个基准测试中平均分为 53.9,高于 Qwen3.5-4B 的 51.1 分。
  • 最强项为工具使用、编码智能体和长上下文检索;最弱项为知识能力。
  • 后训练阶段将 4000 亿 SFT token 与强化学习教师和策略内蒸馏相结合。
  • 预训练、SFT 和 RL 数据集随权重一同发布。

Check out the HF, GitHub repo and Web. Also, feel free to follow us on Twitter and don’t forget to join our 150k+ML SubReddit and Subscribe to our Newsletter. Wait! are you on telegram? now you can join us on telegram as well.

请访问 HF、GitHub 仓库和网站。此外,欢迎在 Twitter 上关注我们,别忘了加入我们有 15 万+成员的 ML SubReddit,并订阅我们的新闻通讯。等等!你在 Telegram 上吗?现在你也可以加入我们的 Telegram 群组。

Need to partner with us for promoting your GitHub Repo OR Hugging Face Page OR Product Release OR Webinar etc.? Connect with us

需要与我们合作推广您的 GitHub 仓库、Hugging Face 页面、产品发布或网络研讨会等吗?请联系我们

The post OpenBMB Releases MiniCPM5-2B: A 2.52B Dense Model Averaging 53.9 Across 34 Benchmarks and Built to Run On Device appeared first on MarkTechPost.

本文最初发布于 MarkTechPost:《OpenBMB 发布 MiniCPM5-2B:一款在 34 个基准测试中平均分达 53.9 的 25.2 亿参数稠密模型,专为设备端运行而设计》

更进一步:量化金融体系

看懂新闻只是起点——沿量化金融路径,把它变成能交付的工程能力

进入量化体系 →

相似阅读

关联信息,但可能不是同一事件