Published 8 September 2026
OpenBMB released MiniCPM5-2B on 7 September 2026, an open model with 2.52 billion parameters that scores an average of 53.9 across the 34 benchmarks in OpenBMB’s own comparison table. Qwen3.5-4B, which carries roughly twice as many parameters, scores 51.1 on that same table. The weights are Apache 2.0 and the training data went out with them.
What is MiniCPM5-2B built for?
MiniCPM5-2B is made to run on hardware you control. Think a laptop, a phone, or one small GPU, rather than a rented data centre. It is the second model in the MiniCPM5 series, after MiniCPM5-1B.
The model is dense, meaning every parameter is used for every word, and it uses the standard LlamaForCausalLM layout. That second point sounds technical but saves you work. vLLM, SGLang, Transformers, llama.cpp, Ollama, LM Studio and MLX load the weights as they are, with no custom code and no patched engine. The context window is 131,072 tokens, roughly a 250-page document in one go.
OpenBMB shipped eight builds at once: the main model, a base checkpoint, a mid-training checkpoint, an SFT-only checkpoint, GGUF and MLX conversions, a 4-bit GPTQ version, and a draft model called MiniCPM5-2B-DSpark for speculative decoding, a trick that speeds up generation without changing what comes out. The training data came too, including UltraData-SFT-Agent-2609 with 500,000 agent samples and UltraData-RL-2609 with more than 80,000 reinforcement learning samples. OpenBMB published JustRL-II-base-model the same day, the starting checkpoint for the reinforcement learning method used here.
MiniCPM5-2B benchmarks against Qwen3.5-4B
The comparison that matters is Qwen3.5-4B, because it is the strongest model in OpenBMB’s own table and it carries about twice the parameters. Read the last row as the honest counterweight to the rest.
| What is measured | MiniCPM5-2B | Qwen3.5-4B |
|---|---|---|
| Average across all 34 benchmarks | 53.9 | 51.1 |
| SWE-bench Verified (fixing real bugs in real code) | 46.4 | 33.6 |
| LiveCodeBench v6 (competitive programming) | 69.1 | 56.4 |
| BFCL v4 (calling the right tool the right way) | 66.6 | 56.8 |
| NoLiMa (finding a fact buried in a long document) | 68.1 | 43.5 |
| MMLU-Pro (broad general knowledge) | 70.8 | 78.0 |
These are OpenBMB’s figures. The model card states that the scores marked with a dagger in its table come from Artificial Analysis and that the rest were reproduced in house, which covers most of the rows above. So the rival numbers were produced by the company making the claim, on its own setup. Treat the direction as informative and the exact gaps as provisional.
Where MiniCPM5-2B falls short
Two gaps are worth knowing before you plan anything around this model.
The first is knowledge. MiniCPM5-2B scores 70.8 on MMLU-Pro against 78.0 for Qwen3.5-4B, and 70.2 on GPQA-Diamond against 77.1. A 2.5 billion parameter model simply holds fewer facts, and no training recipe fixes that. If you want a broad question answering assistant, this is the wrong model.
The second is task length. The tool calling scores are excellent, but they measure short exchanges. On longer agent runs the picture flips: 8.6 on Terminal-Bench v2.1 against 25.8 for Qwen3.5-4B, and 14.4 on SWE-bench Pro against 28.2. MiniCPM5-2B is good at picking the next step and much weaker at holding a plan together across dozens of them.
There is also a number OpenBMB does not publish. The whole pitch is on-device, yet the model card gives no speed and no memory figure for any named device. No tokens per second on a phone, no RAM requirement for the 4-bit build, nothing. For a model sold on where it runs, that is the first thing you need and the one thing the release leaves out.
What this means
If you build agents that call tools or fix code, and they have to run on hardware you control, MiniCPM5-2B is worth testing this week. The team it fits is small and specific: four people running a code review or ticket triage agent on a workstation or an edge box, where the data cannot leave the building. The 46.4 on SWE-bench Verified against 33.6 for a model twice its size is a wide enough gap that it should show up in your own tests, not only in OpenBMB’s table. Start with the 4-bit build, because the speed question the model card ducks is the one that decides whether this ever ships.
For everything else, watch rather than adopt. A broad assistant or a long-running autonomous agent will hit the knowledge ceiling or the planning ceiling within days. The release also fits a pattern from the past few months, visible in Tencent’s EVIE models and in IFM’s K2 Horizon: small open models are closing the gap on narrow, well-defined jobs while barely moving on broad knowledge. What lifts this one above the pattern is what came with it. OpenBMB published the training data and the intermediate checkpoints, so you can measure what each stage of training actually contributed instead of taking the average on trust. That is rarer than a good benchmark score, and it should count for more.
For more information, visit the official announcement of MiniCPM5-2B in the OpenBMB model card.
