MiniCPM5-2B Beats 4B-Class Models on Benchmarks
OpenBMB dropped MiniCPM5-2B and the model card table is the story: average score 53.9 against a comparison set where the best 4B-class model, Qwen3.5-4B, manages 51.1. A 2-billion-parameter model outscoring models twice its size. The gap isn't marginal on the headliner benches either — AIME 2025 86.5 (Qwen3.5-4B: 78.8), LiveCodeBench v6 69.1 (Qwen3.5-4B: 56.4), SWE-bench Verified 46.4 (Qwen3.5-4B: 33.6).
Why It Matters: The Agent Floor Is Collapsing
The number that should stop people is τ²-Bench Telecom: 97.1, with BFCL v4 tool-calling at 66.6 and GAIA Text-103 at 88.7. Those are agentic evaluations. This isn't a chat toy that does homework well — it's a model sized to run on a laptop that can hold tool loops and multi-step tasks. The practical floor for "can I run my agent stack locally on consumer hardware" just moved from ~4B down to ~2B. Add the MLX and GPTQ 4-bit variants OpenBMB shipped the same day, and this runs on Apple Silicon and mid-range GPUs without ceremony.
Read the pattern across our own archive: K2 Horizon 0.9B last week, Puro-2B before that. Small open models are now where the capability-per-parameter action lives, while the frontier chases expensive marginal gains.
The Caveat
The comparison set is chosen by the model's authors, and self-reported tables flatter the self-reported model. GPQA-Diamond (70.2 vs Qwen3.5-4B's 77.1), SWE-bench Pro (14.4 vs 28.2), and Terminal-Bench v2.1 (8.6 vs 25.8) all go to the 4B model — long-horizon agentic work is still its weak spot. The headline "beats 4B class" is true on average and untrue where it's hardest. Independent runs on the Artificial Analysis index would settle it; that's the number I'm watching for this week.
Verdict: real signal, worth a local test — but treat "2B SOTA" as scoped to the authors' table until a third party confirms it.
Frequently Asked Questions
What benchmark scores does MiniCPM5-2B achieve?
OpenBMB reports an average of 53.9 across its comparison set — above Qwen3.5-4B's 51.1 — with AIME 2025 at 86.5, LiveCodeBench v6 at 69.1, SWE-bench Verified at 46.4, τ²-Bench Telecom at 97.1, and BFCL v4 tool-calling at 66.6.
Is MiniCPM5-2B better than 4B-class models?
On average across OpenBMB's comparison set, yes — but 4B models still win on GPQA-Diamond, SWE-bench Pro, and Terminal-Bench v2.1, so the advantage is strongest on math, code, tool-calling, and short agentic tasks rather than long-horizon work.
Can MiniCPM5-2B run locally?
Yes. OpenBMB shipped MLX (Apple Silicon) and GPTQ 4-bit variants alongside the base weights, so it runs on consumer laptops and mid-range GPUs.