MiniCPM5-2B Packs 131K Context and SOTA Results Into 2.5B Parameters

A 2B model beating 4B models

OpenBMB released MiniCPM5-2B this week, and the pitch is the one on-device users always want to hear: smaller model, bigger results. The model is 2.52 billion parameters in BF16, uses the standard LlamaForCausalLM architecture so no custom code or kernels are needed, and carries a 131,072-token context window. On OpenBMB's benchmark suite it averages 53.9, ahead of Qwen3.5-4B at 51.1, IBM's granite-4.2-3B at 42.7, and Liquid AI's LFM2.5-2.6B at 33.2. [1]

Individual results back the average. MiniCPM5-2B scores 97.1 on tau-squared Bench Telecom against 90.4 for LFM2.5-2.6B, 69.1 on LiveCodeBench v6, 86.5 on AIME 2025, 46.4 on SWE-bench Verified, and 70.8 on MMLU-Pro. For a model this size, those are tool-calling and coding numbers that would have needed a much larger model a year ago. [1]

Hybrid thinking and the numbers behind it

The model has a hybrid reasoning mode: a toggle in the chat template enables thinking wrapped in explicit tags, and disabling it injects empty think tags so the same weights serve both fast and deliberate modes. The post-training recipe is unusual in being fully disclosed: training began with 400 billion tokens of deep-thinking SFT, then a three-stage pipeline of SFT, reinforcement learning, and on-policy distillation that OpenBMB says improved reasoning and general capabilities by an average of 10.96 points and agentic skills by 6.96 points. [1]

The company also publishes what worked: 16 expert RL models, five of them agentic, were merged via on-policy distillation using full-vocabulary reverse KL. That level of recipe detail is rare in open releases at any size. [1]

Everything is open, including the data

Apache 2.0 licensing covers the weights, and OpenBMB went further by releasing the UltraData dataset family: web pre-training sets, tiered code data, math data, 500,000 agent SFT samples, and more than 80,000 RL samples. Full checkpoints from base through midtrain, SFT, and the final RL stage are public too, so anyone can reproduce or extend the pipeline rather than starting from scratch. [1]

Running it on your devices

Deployment is where this model earns its spot on a phone. Official support covers vLLM, SGLang, Transformers, llama.cpp, Ollama, LM Studio, MLX, and vLLM Ascend, with ready-made GGUF, MLX 4-bit, GPTQ 4-bit, and LiteRT builds, plus a small draft model for speculative decoding. A FlagOS adaptation runs across nine chip platforms including ARM v9. Tool calling emits XML-style calls that SGLang's built-in parser converts to OpenAI-compatible tool calls, so agent frameworks work without glue code. [1] [2]

For the size class this matters in, the comparison that counts is against what already runs on phones. LFM2.5-2.6B trailed this model by more than 20 points on average in OpenBMB's table, and a 53.9 average from 2.5 billion parameters narrows the gap between what fits on a mid-range device and what needed the cloud last year. [1] [3]