MiniCPM5-2B puts a 2.5B open model above every 4B model in its comparison
OpenBMB shipped MiniCPM5-2B, a dense 2.5-billion-parameter model under Apache-2.0 with a 131,072-token context, posting a 53.9 average that beats every open 4B model it was tested against — and releasing the UltraData training sets behind it. If you run local or on-device AI, this changes the cost-capability tradeoff.
September 7, 2026. OpenBMB released MiniCPM5-2B, the second model in its MiniCPM5 family after MiniCPM5-1B. It is a dense 2.5-billion-parameter transformer under the Apache-2.0 license, with a 131,072-token context window. Its 53.9 average score puts it not only at the top of its own size class, but above every 4B model in the comparison — the best of them tops out at 51.1. Why it matters: the “small but capable” frontier just moved, and OpenBMB published the training data that produced it.
A 2B model that beats 4B models on average
The headline number is the capability-radar average. MiniCPM5-2B scores 53.9, against 51.1 for Qwen3.5-4B, 42.7 for granite-4.2-3B, 32.6 for Nemotron-3-Nano-4B, and 28.4 for LFM2.5-8B-A1B. Inside its own weight class the gap is even wider: Qwen3.5-2B reaches 28.0 and Gemma-4-E2B-it only 24.6.
The inversion is worth pausing on. A model with half the parameter count of Qwen3.5-4B delivers more capability across the measured spectrum. The architecture is otherwise conventional — a standard LlamaForCausalLM with 42 layers, GQA attention using 16 query heads for 2 key-value heads, and 1.98 billion non-embedding parameters.
The gains concentrate exactly where you would not expect them from a small model: math reasoning, code, long context, tool use, and agentic tasks.
The small-model race, in context
MiniCPM5-2B belongs to a broader trend: the compression of the frontier. Two years ago, a 2-billion-parameter model was still a lab curiosity; today it rivals the 4B models that once marked the lower bound of “serious.” Gemma, Phi, Qwen, and LFM have all pushed the same direction, and OpenBMB — the Tsinghua-born team behind MiniCPM — has specialized in this niche since the first generation.
The motivation is not bragging rights. It is economic and practical: a model that fits in under 2 GB once quantized to 4-bit can run on a phone, a single-board computer, or an entry-level GPU, without sending any data to a cloud. For privacy or latency use cases, that is the only acceptable architecture.
OpenBMB’s bet is that the winning small models will be the ones whose training data is open enough to audit — a positioning that separates it from labs that ship compact weights with a closed recipe.
The numbers that decide a deployment
Averages do not close a procurement decision. The raw scores on the benchmarks that matter:
- AIME 2025 and AIME 2026: 86.5 on both — level with far larger models.
- LiveCodeBench v6: 69.1, ahead of granite-4.2-3B (58.9) and Qwen3.5-4B (56.4).
- MATH-500: 94.6.
- MMLU-Pro: 70.8.
- SWE-bench Verified: 46.4 — far above the other 2B models (6.0 for LFM2.5-2.6B).
- NoLiMa (long context): 68.1, where the other 2B models collapse (0.7 for LFM2.5-2.6B) — a level that enables processing whole documents without truncation.
The picture softens on the hardest tasks. On SWE-bench Pro, MiniCPM5-2B posts 14.4 against 28.2 for Qwen3.5-4B; on Terminal-Bench v2.1, 8.6 against 25.8. In other words, for real-environment agentic coding, the 4B models keep the lead. The 2B model’s edge is in the average, not on every extreme.
Perspective still matters against frontier models. On HLE (Humanity’s Last Exam), MiniCPM5-2B scores 8.9, while leading models pass 30. This is not a model that “does everything” — it is one that maximizes capability per byte of memory. The right reading is not “a 2B replaces a frontier model,” but “a 2B replaces a 4B.”
There is also a familiar caveat: math and coding leaderboards are the most aggressively optimized benchmarks in the field, and a small model that tops them is not automatically better at your internal tasks. The scores are a useful signal of the architecture’s ceiling, not a guarantee of fit.
What “on-device” actually changes
Running an LLM locally is not just “it works without a connection.” Three properties follow directly:
- Privacy. Sensitive documents — contracts, health data, internal code — never leave the machine.
- Latency. No network round trip: generation starts immediately, decisive for an interactive coding agent.
- Cost. Once the hardware is amortized, the marginal cost per token is zero, where an API call stays billed per million tokens.
MiniCPM5-2B checks all three boxes while publishing its data and weights under Apache-2.0 — meaning a team can audit, fine-tune, and redistribute without legal negotiation.
The real story is the data, not just the weights
What separates this release is what ships alongside it. OpenBMB published, under the UltraData family, the datasets used to train the model:
- UltraX, the high-quality web pre-training corpus;
- UltraData-Code, with L0–L3 tiered code-data management that explains the coding jump;
- UltraData-SFT-Agent-2609, 500,000 agent-training samples;
- UltraData-RL-2609, 80,000-plus reinforcement-learning tasks spanning math, code, general knowledge, and long-context reasoning.
In an ecosystem where most labs keep their recipes secret, releasing both the weights and the data under Apache-2.0 changes what you can audit and reproduce. It is an argument for trust, not just a technical one.
For regulated teams, the difference is concrete: with the data and the license in hand, you can document what went into the model and fine-tune on your own corpus without a commercial agreement — a path that closed models simply do not offer.
A full deployment ecosystem
The model ships for every mainstream runtime: BF16 (the final RL + OPD post-trained release), GGUF for llama.cpp, Ollama, and LM Studio, MLX for Apple Silicon, GPTQ 4-bit, and even a DSpark draft model to accelerate inference. You can run it on a laptop, a phone, or a local cluster with no cloud dependency.
The DSpark draft model deserves a note: it serves speculative decoding, a technique where a small model generates quick drafts that the main model then verifies. OpenBMB ships a dedicated draft rather than forcing teams to train their own, which simplifies rolling out accelerated inference.
Getting started is deliberately low-friction: with Ollama, a ollama run minicpm5 one-liner is enough to get a local chat or coding loop going once the model is pulled, and the MLX build gives Apple Silicon users a native path. The absence of an API dependency means the model can sit behind your own service boundary from day one.
Verdict
If you run local, embedded, or private AI — an on-device assistant, an offline coding agent, sensitive document processing — MiniCPM5-2B is now the reference point of the sub-4B class: near-4B capability, an Apache-2.0 license, published data, and builds for every runtime. If your workload demands heavy agentic coding — SWE-bench Pro, terminal environments — keep a 4B-or-larger model, because that is where the gap still opens. Either way, test on your workload before switching: the 2B model’s impressive average is no substitute for measuring against your own domain.
The UltraData release, finally, has value beyond this model: it gives the community a reproducible basis to study what separates an average 2B from an excellent one, and a clean starting point for fine-tuning — a rare resource in an ecosystem where most labs keep their recipes closed.