MiniCPM5-2B

MiniCPM5-2B

Quick answer: MiniCPM5-2B is OpenBMB's second model in the MiniCPM5 series — a dense 2B-parameter Transformer built for on-device and resource-constrained deployment. It reaches 2B-class open-source SOTA (average 53.9 across OpenBMB's comparison suite) and beats several larger 4B-class models on coding, math, tool use, and agentic tasks. Apache 2.0, 131K context, released September 2026.

At a Glance

Where MiniCPM5-2B leads

• 2B-class open-source SOTA across coding, math, tool use, long-context, and agentic benchmarks per OpenBMB's comparison suite
• Strong math reasoning for its size: 86.5% AIME 2026, 63.8% HMMT Feb 2026
• Native 131K context with reported long-context retrieval scores (59.0% AA-LCR) well above same-size peers

Where it lags

• SWE-bench Pro (14.4%) and Terminal-Bench v2.1 (8.6%) remain far below larger models — coding-agent tasks are still hard at this size
• HLE score (8.9%) reflects the ceiling of a 2B model on frontier knowledge questions

Best for: local/on-device assistants, coding agents, and tool-use workflows where a small footprint matters more than frontier-model ceiling performance.

What MiniCPM5-2B Is

MiniCPM5-2B is the second release in OpenBMB's MiniCPM5 family, following MiniCPM5-1B. It's a dense, standard LlamaForCausalLM architecture (2.52B total parameters, 42 layers, GQA with 16 query / 2 KV heads) trained with the same recipe as MiniCPM5-1B scaled up, plus a full "UltraData" tiered training pipeline (base → mid-train → SFT → RL → On-Policy Distillation).

It ships alongside its full training data stack (UltraX pretraining corpus, UltraData-Code, UltraData-SFT-Agent-2609 with 500K agent samples, UltraData-RL-2609 with 80K+ RL samples) — one of the more transparent open releases in the small-model space. The RL + On-Policy Distillation stage is reported to add +10.96 points average on reasoning/general tasks and +6.96 on agentic tasks over the SFT-only checkpoint.

Specifications

Pricing

Open weights — free to download and self-host; cost is inference/hosting only. No hosted API pricing published by OpenBMB at release.

Context Window

MiniCPM5-2B has a 131,072-token context window — roughly 250+ pages of text in a single request. OpenBMB reports 59.0% on AA-LCR (long-context retrieval), notably ahead of same-size peers (Qwen3.5-2B: 5.3%, Gemma-4-E2B-it: 28.7%), suggesting the long-context claim is backed by real retrieval performance rather than just an extended position-embedding ceiling.

Public Benchmark Scores

Scores below are self-reported by OpenBMB in the MiniCPM5-2B model card, evaluated against a comparison set of same-size and adjacent-size open models. They are not Benchgen measurements. Figures marked † in the original card come from the official Artificial Analysis release; all others were reproduced internally by OpenBMB.

Not mapped (excluded, no matching Benchgen benchmark entity): LCB-Pro 25Q2 (Easy/Medium), SciCode (wbg) — a variant reported alongside a different weighting than our SciCode entity, NoLiMa, LongBenchPro, GAIA Text-103, ClawGym-Bench, WildClaw, QwenClaw. These are excluded rather than force-mapped to a similarly-named existing entity, since methodology details (task counts, exact variant) were not independently confirmed to match. Flagged for a future dedicated benchmark-config pass if cross-model demand emerges.

MiniCPM5-2B vs Alternatives

Rival scores are as reported in OpenBMB's own comparison table (same source as MiniCPM5-2B's scores above), not independently re-verified by Benchgen.

MiniCPM5-2B's pitch is that a well-executed dense 2B model with strong post-training (RL + OPD) can match or beat sloppier 4B-class releases on real capability signals like tool use and long-context retrieval — worth checking against your own workload before assuming bigger always wins on-device.

How MiniCPM5-2B Performs on Real Agent Tasks

MiniCPM5-2B's headline story is agentic capability at a small footprint: BFCL v4 (66.6%) and τ²-Bench Telecom (97.1%) both lead its comparison set, and the model's SFT data explicitly includes 500K agent-training samples (UltraData-SFT-Agent-2609). That said, coding-agent benchmarks that require sustained multi-turn execution (SWE-bench Pro 14.4%, Terminal-Bench 8.6%) show the ceiling is still well below larger models — tool-calling accuracy and long-horizon task completion are not the same capability, and builders evaluating this model for an agent harness should test both independently rather than extrapolating from one.

Related Models & Benchmarks

• Compare: MiniCPM5-1B · Qwen3.5-2B
• Benchmarks: GPQA Diamond · AIME 2026 · BFCL v4

Frequently Asked Questions

Specs and scores sourced from OpenBMB's official MiniCPM5-2B model card. Third-party comparison scores attributed inline. Last updated 2026-09-08.