OpenBMB Releases MiniCPM5-2B: 2.5B-Parameter On-Device LLM Outperforming Larger Models on Tool Use and Coding
Summary
- • OpenBMB releases MiniCPM5-2B: 2.52B params, 131K context window, Apache 2.0 license
- • Averages 53.9 across 34 benchmarks — outperforming the larger Qwen3.5-4B (51.1)
- • Strongest on tool use (97.1 τ²-Bench Telecom) and coding agents; weaker on general knowledge
- • Full training datasets and intermediate checkpoints released; all training stages verifiable
Details
MiniCPM5-2B: 2.52B params, 42 layers, GQA (16Q/2KV heads), 131K native context
Second checkpoint in the MiniCPM5 series; standard LlamaForCausalLM architecture with 1.98B non-embedding parameters, deployable with no custom kernels across mainstream inference engines.
53.9 average across 34 benchmarks — beats Qwen3.5-4B (51.1), a larger model
OpenBMB compared against LFM2.5-2.6B, Qwen3.5-2B, and Gemma-4-E2B-it in class; also listed Qwen3.5-4B (51.1), granite-4.2-3B (42.7), and LFM2.5-8B-A1B as reference baselines.
Tool use: 97.1 τ²-Bench Telecom vs 6.8; BFCL v4 66.6; τ³-Bench Banking 20.8 vs 6.8
Tool-calling benchmarks show the widest performance margins over any comparator, positioning MiniCPM5-2B as the standout choice for agentic and API-calling workloads at the 2-3B scale.
Code: 69.1 LiveCodeBench v6 vs 56.4; SWE-bench Verified 46.4 vs 33.6
Coding agent benchmarks show significant leads over the next best competitor, making it viable for code-completion agents and software engineering tasks.
OPD merges 16 RL experts (5 agentic) via reverse KL divergence advantage estimation
On-policy distillation replaces the standard verification-based advantage with full-vocabulary reverse KL divergence between student and teacher logits; reuses RL prompts as distillation data so no new corpus is needed.
Open datasets: SFT Agent (500K samples), RL (80K+ samples), plus 5 other corpora
Released alongside weights: Ultra-FineWeb, Ultra-FineWeb-L3, UltraX, UltraData-Code, UltraData-Math, UltraData-SFT-2605, UltraData-SFT-Agent-2609, UltraData-RL-2609. Intermediate checkpoints (Base, Midtrain, SFT-only) also published.
Knowledge gap: 70.8 MMLU-Pro vs 78.0; trails on HLE and GPQA-Diamond
MiniCPM5-2B is explicitly a specialist, not a generalist — trails larger models on knowledge-heavy benchmarks, positioning it for agentic/tool/code use cases rather than broad QA.
Runs on Ollama, llama.cpp, LM Studio, MLX, vLLM, SGLang, Transformers, FlagOS
Apache 2.0 license and standard Llama architecture mean developers can drop it into existing local or cloud inference stacks immediately without code changes.
Source: OpenBMB release writeup via Grok Sweep aggregator, September 7, 2026.
What This Means
MiniCPM5-2B is a compelling on-device option for agentic and tool-calling use cases, achieving benchmark results that beat some 4B+ models while staying within a 2.5B footprint — critical for edge deployment and cost-sensitive inference. Its broad compatibility with Ollama, llama.cpp, and LM Studio means developers can deploy it immediately in existing local workflows. The full release of training data and intermediate checkpoints is unusually transparent by open-source standards, making the on-policy distillation claims checkable — which matters more than headline averages. Teams building coding agents or tool-calling pipelines should evaluate it as a serious size-efficient alternative.
Sources
- OpenBMB Releases MiniCPM5-2BMarktechpost