OpenBMB MiniCPM5-2B leads open-weight models below 4B parameters on Intelligence Index v4.2

OpenBMB MiniCPM5-2B leads open-weight models below 4B parameters on Intelligence Index v4.2

OpenBMB's MiniCPM5-2B scored 15 on the Artificial Analysis Intelligence Index v4.2, the highest result among open-weight models with fewer than 4B total parameters.

MiniCPM5-2B is a 2.6B-parameter dense reasoning model with text input and output. It is released under the Apache 2.0 license.

Its score is one point below Ling 3.0 Tiny, which scored 16 with approximately three times as many total parameters. The next-highest score among open-weight models below 4B parameters is Granite 4.2 3B, at 11.

Key results

• Highest score among open-weight models below 4B parameters: MiniCPM5-2B's score of 15 is four points ahead of Granite 4.2 3B. With 2.6B total parameters, it is one point ahead of the estimated 14 for Qwen3.5 4B (Reasoning), while using 44% fewer parameters. It matches the estimated score of Qwen3.5 9B (Reasoning), which has approximately four times as many parameters. Because MiniCPM5-2B is a dense model, its size advantage primarily affects memory footprint rather than active-parameter compute.
• Strong agentic performance: Its GDPval-AA v2 Elo score of 831 leads models below 4B parameters. On τ³-Banking, it is joint-first with Ling 3.0 Tiny at 21%, compared with 8% for the next-best model, Granite 4.2 8B. On AA-Briefcase, MiniCPM5-2B ranked second among the measured models in the comparison set, with an Elo score of 438. It was ahead of Granite 4.2 8B at 324 and behind Ling 3.0 Tiny at 485.
• Weaker results in knowledge, coding, and long-context tasks: MiniCPM5-2B ranked seventh in the comparison set on Humanity's Last Exam, scoring 9% compared with 16% for Gemma 4 12B (Reasoning). It ranked eighth on Terminal-Bench v2.1 at 9%, behind Qwen3.5 9B (Reasoning) at 29%, and scored 0% on CritPt. On SciCode, it ranked second among five measured models at 26%, behind Granite 4.2 8B at 31%. On AA-LCR v1.1, it scored 59%, placing fifth and one point behind Ling 3.0 Tiny at 60%. On GDP.pdf, a professional document-reasoning evaluation, it passed 1% of tasks outright, compared with 2% for gpt-oss-20b (high).
• Low AA-Omniscience score driven by abstention: MiniCPM5-2B scored -12 on AA-Omniscience after attempting only 29% of questions. Its Non-Hallucination Rate was 78%, while its accuracy was 8%. That accuracy was one point below Ling 3.0 Tiny at 9% and half the 16% recorded by Qwen3.5 9B (Reasoning). Models that attempted substantially more questions received lower overall scores, including Qwen3.5 9B (Reasoning) at -53 and gpt-oss-20b (high) at -63.
• Efficient output-token usage: MiniCPM5-2B generated 19k output tokens per Intelligence Index task, joint-lowest in the comparison set with Granite 4.2 3B. Ling 3.0 Tiny used 56k output tokens, approximately three times as many, for one additional index point.

Model details

• Parameters: 2.6B, dense
• Context window: 131k tokens
• Input modalities: Text only
• License: Apache 2.0

Agentic performance

Agentic capability is MiniCPM5-2B's strongest area at this parameter scale. On AA-Briefcase, the agentic knowledge-work evaluation, it achieved an Elo score of 438. That placed it second in the comparison set behind Ling 3.0 Tiny at 485 and ahead of Granite 4.2 8B at 324.

The Agentic Index is a weighted average of three agentic evaluations included in the Intelligence Index: AA-Briefcase, GDPval-AA v2, and τ³-Banking.

On GDPval-AA v2, which evaluates real-world work tasks against a human baseline of 1,000, MiniCPM5-2B reached an Elo score of 831. This was approximately 110 points ahead of Ling 3.0 Tiny at 718 and approximately 180 points ahead of Granite 4.2 8B at 647. Other models of a similar size scored lower, including LFM2.5-2.6B at 204 and Gemma 4 E4B (Reasoning) at 178.

Token efficiency

MiniCPM5-2B used 19k output tokens per Artificial Analysis Intelligence Index task, including 11k reasoning tokens. Ling 3.0 Tiny used 56k output tokens, while Granite 4.2 8B used 33k. Output-token usage is relevant to the on-device and edge deployments targeted by a 2.6B-parameter model.

Results across evaluations

MiniCPM5-2B scored 23 on the earlier Artificial Analysis Intelligence Index v4.1.1. Its score of 15 on v4.2 reflects changes to the evaluation mix and weightings, so scores from the two versions are not directly comparable.

The model weights are available under the Apache 2.0 license at Hugging Face.