MiniCPM5-1B is a 1-billion-parameter model built for on-device and resource-constrained scenarios. Ultra-FineWeb, Ultra-FineWeb-L3, UltraData-Math, and UltraData-SFT-2605 have been open-sourced.
Highlights
18.2 Artificial Analysis Index 4.0 — Benchmark results against same-size open-source models：Comparison results against same-size open-source models are given in the evaluation table, covering agent tool calling, code generation, and complex reasoning.
Excellent Overall Text Performance Ranking：On the Artificial Analysis (AA) Intelligence Index, MiniCPM5-1B (thinking) scores 18.2 and MiniCPM5-1B (non-thinking) scores 17.9; Qwen3.5-2B, shown on the same page, scores 16.3.
Desktop Pet：After INT4 quantization the weights are about 0.5GB and can run in on-device environments such as phones and browsers.
At the 1B parameter scale it can drive on-device applications, for example a "desktop pet" running locally.
Technical highlights
Three-stage training (base → mid → post) + fully open-source corpora：Model training spans three stages: base training → mid training → post-training. Base training uses stable training plus decay training to build core language capability and training stability, while mid training strengthens target capabilities and adapts to the target data distribution. The supporting corpora are fully open-source, including Ultra-FineWeb, Ultra-FineWeb-L3, UltraData-Math, and the post-training corpus UltraData-SFT-2605.
SFT → RL → OPD in three steps：Post-training follows three steps: SFT → RL → OPD. First, 200B tokens of deep-thinking SFT and 200B tokens of hybrid-thinking SFT build deep-thinking, hybrid-thinking, and general dialogue capabilities. Then specialized RL teacher models are trained for domains such as mathematics, code, closed-book QA, and writing. Finally, On-Policy Distillation (OPD) distills the capabilities of multiple teachers back into one released model.
Result: average score on math / code / instruction following +16 points; the share of responses hitting the maximum token budget -29 percentage points.
Reasoning RL + OPD Improvements：Reasoning RL: based on DAPO-Math-17k and inspired by the minimalist JustRL recipe, it uses two-stage length scheduling to improve reasoning accuracy while shortening overly long responses. It also incorporates TriviaQA, NQ-Open, LongWriter-Zero-RLData, synthetic RLVR, and pairwise RLHF signals.
OPD improvements: within the RL framework, reverse KL divergence is used as the advantage estimate, replacing the original validation-based advantage calculation. RKL is computed over the union of student and teacher top-k logits, balancing accuracy and efficiency. It also reuses each RL teacher's in-domain prompts as distillation data, reducing additional data-curation steps.
Desktop Pet：We also provide OpenBMB/MiniCPM-Desk-Pet, a desktop pet powered locally by MiniCPM5-1B. It supports multiple runtime paths on Apple Silicon, NVIDIA GPU, and CPU; can work together with coding agents such as Cursor, Claude Code, and Codex; and supports LoRA-based character switching.
Benchmarks
Overall：The comparison models are LFM2.5-1.2B-Thinking, Qwen3-0.6B/think, and Qwen3.5-0.8B/think; per-task scores are given in the table below, covering tool calling, code generation, and complex reasoning. Suitable for local code agents, tool assistants, and reasoning assistants.