One-line: 2.5B dense follow-up to MiniCPM5-1B (OpenBMB/Tsinghua NLP + ModelBest) — AA Intelligence Index v4.2 score of 15, the highest of any open-weight model under 4B total params (v4.1.1: 23), with class-leading agentic/coding numbers (SWE-bench Verified 46.4, LiveCodeBench v6 69.1, BFCL v4 66.6) at a size that runs fully on-device. Released ~2026-09-07 (repo created 09-06; launch coverage 09-07), surfaced via the trending window (lastModified 09-12).
Coding benchmarks
Benchmark
Score
Source
HumanEval
—
not published
HumanEval+
—
not published
MBPP
—
not published
LiveCodeBench
69.1 (v6)
lab (internally reproduced)
LCB-Pro 25Q2 Easy / Medium
68.0 / 17.5
lab
OJBench
32.5
lab
SciCode (wbg)
26.3
Artificial Analysis (†) — 2nd of the 5-model set, behind Granite 4.2 8B (31%) per AA article
Artificial Analysis (independent) — highest of any open model under 4B total params
MMLU-Pro
70.8
lab
GPQA Diamond
70.2
Artificial Analysis (†)
MATH-500
94.6
lab
AIME 2025 / AIME 2026
86.5 / 86.5
lab
HMMT Feb 2026
63.8
lab
Humanity’s Last Exam
8.9
Artificial Analysis (†) — 7th in the measured set per AA article
Efficiency
Active params: 2.52B dense (no MoE, all active; 1.98B non-embedding)
Context length: 131,072 tokens (128K native)
VRAM (fp16 rough): ~5.0 GB weights (2.52B × 2 bytes); official GGUF repo (openbmb/MiniCPM5-2B-GGUF) with 4-bit ≈ ~1.5 GB; MLX 4-bit for Apple Silicon
Runs on: vLLM, SGLang, Transformers, llama.cpp (GGUF), Ollama, LM Studio, MLX, FlagOS — standard LlamaForCausalLM, no custom kernels needed
What makes it notable
MiniCPM5-2B is the strongest open-weight model at ≤2.5B params across both coding and agentic suites: 46.4 SWE-bench Verified is 12.8 pts over Qwen3.5-4B (33.6) — the first sub-3B model with meaningful repo-level agent coding — and BFCL v4 66.6 leads the 2B (and most 4B) class. AA’s independent Index agrees: 15 on v4.2 is the best under 4B total params, with clear agentic signal (GDPval-AA Elo 831, AA-LCR 59) and 128K native context. Same hybrid think/no-think, RL + on-policy-distillation UltraData recipe as MiniCPM5-1B but with a much stronger agentic profile — the natural default for local tool-using agents on laptops/SBC-class hardware. Caveats: per-benchmark numbers are lab-reproduced unless † (AA-sourced); Aider polyglot and τ-Bench main set are unpublished, and Text-103 GAIA-type results are lab-only. SWE-bench Verified 46.4 at 2.5B should be treated as lab-reported until third-party reproduction.