AI News AI资讯 1h ago Updated 1h ago 更新于 1小时前 48

OpenBMB Releases MiniCPM5-2B: A 2.52B Dense Model Averaging 53.9 Across 34 Benchmarks and Built to Run On Device OpenBMB发布MiniCPM5-2B:25.2亿参数密集模型,34项基准测试平均53.9分,专为端侧运行打造

MiniCPM5-2B is a 2.52B dense causal language model with a 131,072-token context window, averaging 53.9 across 34 benchmarks, outperforming same-size peers and even larger models like Qwen3.5-4B (51.1) The model excels in tool use (97.1 on τ²-Bench Telecom, 66.6 on BFCL v4), code reasoning (69.1 on LiveCodeBench v6, 46.4 on SWE-bench Verified), and long-context retrieval (68.1 on NoLiMa), but trails on general knowledge benchmarks Training pipeline combines 400B tokens of deep-thinking SFT, speci MiniCPM5-2B是2.52B参数的密集因果语言模型,34个基准测试平均53.9分,超越同级别Qwen3.5-4B(51.1分) 采用标准LlamaForCausalLM架构,131,072 token原生上下文窗口,Apache 2.0开源许可,兼容vLLM、SGLang、llama.cpp等主流推理引擎 训练流程为400B token深度思考SFT → JustRL II算法RL训练 → 在线策略蒸馏(OPD),融合16个RL专家(含5个代理专家) 优势领域:工具使用(τ²-Bench Telecom 97.1)、代码推理(LiveCodeBench v6 69.1)、长上下文检索(N

68
Hot 热度
65
Quality 质量
70
Impact 影响力

Analysis 深度分析

TL;DR

  • MiniCPM5-2B is a 2.52B dense causal language model with a 131,072-token context window, averaging 53.9 across 34 benchmarks, outperforming same-size peers and even larger models like Qwen3.5-4B (51.1)
  • The model excels in tool use (97.1 on τ²-Bench Telecom, 66.6 on BFCL v4), code reasoning (69.1 on LiveCodeBench v6, 46.4 on SWE-bench Verified), and long-context retrieval (68.1 on NoLiMa), but trails on general knowledge benchmarks
  • Training pipeline combines 400B tokens of deep-thinking SFT, specialized RL teachers using the critic-based JustRL II algorithm, and on-policy distillation (OPD) merging 16 RL experts into a single model
  • OpenBMB released the full training data (Ultra-FineWeb, UltraData series) and intermediate checkpoints (Base, Midtrain, SFT-only), enabling independent verification of the RL+OPD claims
  • Apache 2.0 licensed with standard LlamaForCausalLM architecture, ensuring compatibility with mainstream inference engines (vLLM, SGLang, llama.cpp, Ollama, MLX) without custom kernels

Why It Matters

MiniCPM5-2B demonstrates that a sub-3B parameter model can compete with significantly larger counterparts on agentic and tool-calling workloads, making it a viable on-device option for resource-constrained deployments. The open release of training data and intermediate checkpoints sets a new transparency standard, allowing researchers to audit and replicate the post-training pipeline rather than relying solely on headline benchmark averages.

Technical Details

  • Architecture: 42-layer dense transformer with grouped-query attention (16 query heads, 2 key/value heads), 131,072-token native context window, standard LlamaForCausalLM architecture for plug-and-play compatibility
  • Benchmark Performance: Averages 53.9 across 34 benchmarks; leads in tool use (τ²-Bench Telecom: 97.1, BFCL v4: 66.6), coding (LiveCodeBench v6: 69.1, SWE-bench Verified: 46.4), and long-context retrieval (NoLiMa: 68.1); weaker on knowledge-heavy tasks (MMLU-Pro: 70.8 vs. 78.0 for larger models)
  • Training Pipeline: UltraData tiered data management with base training (stable + decay phases), mid-training adaptation, 400B tokens of deep-thinking SFT, specialized RL teachers for math/code/agentic/writing using JustRL II critic-based algorithm, and on-policy distillation merging 16 RL experts
  • On-Policy Distillation: Replaces verification-based advantage with full-vocabulary reverse KL divergence between student and teacher logits at each response position, reusing RL prompts as distillation data without building new corpora
  • Open Ecosystem: Apache 2.0 weights, full dataset release (Ultra-FineWeb, Ultra-FineWeb-L3, UltraX, UltraData-Code, UltraData-Math, UltraData-SFT-2605, UltraData-SFT-Agent-2609 with 500K agent samples, UltraData-RL-2609 with 80K+ RL samples), and intermediate checkpoints for reproducibility

Industry Insight

The success of on-policy distillation in merging multiple RL experts into a compact 2.52B model suggests a scalable path toward high-performance edge AI, reducing dependency on massive parameter counts for agentic workloads. The open data and checkpoint release model could accelerate community-driven improvements and audits, raising the bar for transparency in the competitive small-language-model race. Practitioners should prioritize MiniCPM5-2B for tool-calling and coding agent deployments on edge devices, while supplementing with larger models for knowledge-intensive tasks.

TL;DR

  • MiniCPM5-2B是2.52B参数的密集因果语言模型,34个基准测试平均53.9分,超越同级别Qwen3.5-4B(51.1分)
  • 采用标准LlamaForCausalLM架构,131,072 token原生上下文窗口,Apache 2.0开源许可,兼容vLLM、SGLang、llama.cpp等主流推理引擎
  • 训练流程为400B token深度思考SFT → JustRL II算法RL训练 → 在线策略蒸馏(OPD),融合16个RL专家(含5个代理专家)
  • 优势领域:工具使用(τ²-Bench Telecom 97.1)、代码推理(LiveCodeBench v6 69.1)、长上下文检索(NoLiMa 68.1);通用知识相对较弱(MMLU-Pro 70.8)
  • 完整开源训练数据(Ultra-FineWeb、UltraData系列)及中间检查点(Base、Midtrain、SFT-only),支持逐阶段复现验证

为什么值得看

MiniCPM5-2B为端侧Agent和工具调用场景提供了经过验证的轻量级方案,其工具使用与代码推理能力在同尺寸模型中领先。完整的开源数据与中间检查点使RL+OPD训练流程可复现,对研究小模型强化学习路径具有参考价值。

技术解析

  • 架构规格:2,516,756,480参数(1,981,982,720非嵌入参数),42层,分组查询注意力(16 query heads / 2 KV heads),标准LlamaForCausalLM架构,无需自定义kernel即可部署
  • 训练流程:Base训练(稳定+衰减阶段)→ Midtrain适配目标分布 → 400B token深度思考SFT → 多领域RL教师训练(数学、代码、代理、写作,使用JustRL II critic-based算法)→ 在线策略蒸馏(OPD)
  • OPD技术:在响应位置计算学生与教师logits的全词表反向KL散度作为优势估计,替代传统验证-based优势;复用RL提示作为蒸馏数据,无需构建新语料
  • 基准表现:工具使用领先显著(BFCL v4 66.6、τ²-Bench Telecom 97.1);代码代理(SWE-bench Verified 46.4);长上下文NoLiMa 68.1但LongBench v2 43.7;通用知识落后更大模型(MMLU-Pro 70.8 vs 78.0)
  • 开源资产:Ultra-FineWeb、Ultra-FineWeb-L3、UltraX、UltraData-Code、UltraData-Math、UltraData-SFT-2605、UltraData-SFT-Agent-2609(500K代理样本)、UltraData-RL-2609(80K+ RL样本)及Base/Midtrain/SFT-only中间检查点

行业启示

  • 小模型(<3B)在特定垂直场景(工具调用、代码代理)可通过RL+蒸馏策略达到可用水平,端侧部署成本显著降低,适合资源受限的Agent应用
  • OPD技术用反向KL散度替代验证-based优势估计,为多专家模型融合提供了无需额外数据的高效路径,值得在后续小模型训练中探索
  • 完整开源训练数据与中间检查点成为可复现研究的标杆,行业应重视训练链路的透明度以推动小模型技术迭代

Disclaimer: The above content is generated by AI and is for reference only. 免责声明:以上内容由 AI 生成,仅供参考。

Open Source 开源 LLM 大模型 Benchmark 基准测试 Product Launch 产品发布 Inference 推理