A 2.52B model outscores 4B rivals, on a benchmark set its own maker ran

OpenBMB's own benchmark table for MiniCPM5-2B lists things a 2.5-billion-parameter checkpoint is not supposed to be good at. It scores 69.1 on LiveCodeBench v6, 46.4 on SWE-bench Verified and 66.6 on BFCL v4, the function-calling test. The model itself carries about 2.52 billion parameters, of which roughly 1.98 billion are non-embedding.

That footprint is the point. OpenBMB, the open-source arm associated with ModelBest, is shipping MiniCPM5-2B as something a phone, a laptop or an edge box can host rather than something a GPU cluster serves. It is the same bet Liquid AI's LFM2.5-2.6B has already placed for agentic work. Whether the published margins survive contact with independent testing is the open question, and the release materials make that question easy to ask.

What OpenBMB actually shipped

MiniCPM5-2B is a dense model rather than a mixture of experts, and it lands under Apache-2.0. It is the second drop in the MiniCPM5 line, after MiniCPM5-1B. Native context runs to 131,072 tokens.

The architecture is deliberately unremarkable, which counts as a feature at this size. It uses a standard LlamaForCausalLM layout: 42 layers with grouped-query attention, 16 query heads and 2 key/value heads. Because nothing in the graph is custom, mainstream inference engines can load the weights without a patched model definition or a hand-written kernel.

The 53.9 average and who ran the test

Across OpenBMB's 34-benchmark comparison set, MiniCPM5-2B averages 53.9. That puts it ahead of several larger open baselines in the same table, including Qwen3.5-4B at 51.1 and granite-4.2-3B at 42.7. Same-size peers such as LFM2.5-2.6B and Qwen3.5-2B sit further back. The field of small, competitive open models keeps widening, as the recent sub-200M surge suggests.

Provenance is the part worth slowing down for. OpenBMB separates rows sourced from Artificial Analysis from rows it reproduced internally, so readers can tell the two kinds of number apart. A blended average is not the same claim as an all-external one, and 53.9 should be read against that split rather than above it.

Figures as published by OpenBMB, drawn from vendor-reproduced and Artificial Analysis rows.

Where a 2B model beats 4B, and where it does not

The wins cluster in agentic work. LiveCodeBench v6 runs 69.1 against 56.4 for the 4B Qwen reference, SWE-bench Verified 46.4 against 33.6, and NoLiMa's long-context retrieval 68.1 against 43.5. OpenBMB also reports τ²-Bench Telecom at 97.1 and BFCL v4 at 66.6.

Knowledge-heavy rows tell the other half of the story. MMLU-Pro lands at 70.8 for MiniCPM5-2B against 78.0 for the larger Qwen model, a gap running the opposite direction. OpenBMB's positioning matches the data: a compact agent and coding companion rather than a general-knowledge giant. Beating larger models on narrow tasks rather than broad knowledge is a pattern the open-weight field keeps repeating, from Shieldstral's 3B safety classifier to sub-billion agent builds.

The 16 experts behind the jump

Post-training runs through a pipeline OpenBMB calls UltraData tiered management. It starts with deep-thinking supervised fine-tuning on about 400 billion tokens. Specialized reinforcement-learning teachers for math, code, agents and writing follow, trained with a critic-based algorithm the lab calls JustRL II. The final step is on-policy distillation that folds 16 RL experts, five of them agentic, into the single student that ships.

OpenBMB attributes average gains of about 10.96 points on reasoning and general suites, and 6.96 points on agentic suites, to that RL and distillation stage. Rather than ask readers to take the claim on faith, it published the intermediate Base, Midtrain and SFT-only checkpoints, which lets anyone measure each contribution on its own. Distillation is the stage that compresses 16 experts into one student, and it is a step the research literature keeps reworking, as work on evolving distillation contexts shows.

What you can actually run it on

Runtime coverage spans vLLM, SGLang, Transformers, llama.cpp, Ollama, LM Studio, MLX, LiteRT and FlagOS multi-chip builds. GGUF, MLX and GPTQ packages target local assistants that need tool calling and long context without a datacenter GPU. The UltraData corpora used in training, covering web, code, math, and agent SFT and RL samples, are published too.

Weights, data, intermediate checkpoints and quantized builds in a single release lower the cost of checking the work. That kind of open packaging is still rare enough to state plainly. Running capable models on local hardware is a case others have made too, including LFM2.5-Encoders on CPU.

The 53.9 average will get quoted. The number that decides whether MiniCPM5-2B matters is narrower: how much of the coding and tool-calling lead over 4B open baselines survives when someone else runs the harness. OpenBMB has handed over what is needed to find out.