OpenBMB's MiniCPM5-2B Tops Sub-4B Charts While Burning 3x Fewer Tokens

OpenBMB's MiniCPM5-2B lands at the top of the sub-4B open-weights charts on Artificial Analysis's Intelligence Index, scoring 15 points against a median of 1 for comparable models. That puts it four points ahead of Granite 4.2 3B (11) and level with an estimated Qwen3.5 9B reasoning variant at roughly four times its size. The only sub-4B model above it is Ling 3.0 Tiny at 16, which carries about 3x the total parameters. At 2.6B dense parameters, it also uses fewer output tokens per task than almost anything else in its class.

Specs at a glance

• 2.6B total parameters, dense (all parameters active at inference)
• 131k token context window
• Text in, text out, reasoning-style chain-of-thought
• Apache 2.0 license, weights on Hugging Face
• Knowledge cutoff: December 2025

Built for agents, not just benchmarks

On AA-Briefcase, which scores agents on knowledge work tasks, MiniCPM5-2B posts an Elo of 438, second in its comparison set behind Ling 3.0 Tiny (485) and well ahead of the larger Granite 4.2 8B (324). On the tau-cubed Banking tool-use benchmark, it ties Ling 3.0 Tiny for first at 21%, while the next best model sits at 8%.

The GDPval-AA v2 result is the most striking. This test measures real-world professional work tasks against a human baseline of 1,000 Elo. MiniCPM5-2B reaches 831, about 110 points ahead of Ling 3.0 Tiny and 180 ahead of Granite 4.2 8B. Other 2-3B models fall far behind: LFM2.5-2.6B scores 204, and Gemma 4 E4B (Reasoning) scores 178.

Unusually low token burn

Reasoning models routinely exhaust output budgets. MiniCPM5-2B uses 19k output tokens per Intelligence Index task, roughly 11k of them reasoning tokens, tied for the lowest in its comparison set. Ling 3.0 Tiny spends 56k tokens to earn one extra index point; Granite 4.2 8B spends 33k. For edge or on-device deployments where tokens translate directly to latency and battery drain, that gap carries more weight than any single benchmark number.

Where it gives ground

The weaknesses cluster around raw knowledge recall and hard coding:

• Humanity's Last Exam: 9%, 7th in its set, behind Gemma 4 12B (Reasoning) at 16%
• Terminal-Bench v2.1: 9%, 8th, well behind Qwen3.5 9B (Reasoning) at 29%
• CritPt (physics reasoning): 0%
• GDP.pdf, a professional document reasoning eval: passes 1% of tasks
• AA-LCR v1.1 long-context reasoning: 59%, one point behind Ling 3.0 Tiny

The AA-Omniscience score of -12 deserves context. MiniCPM5-2B attempts only 29% of questions, giving it a 78% Non-Hallucination Rate and raw accuracy of 8%. Peers that attempt more questions get penalized harder: Qwen3.5 9B (Reasoning) scores -53, and gpt-oss-20b (high) scores -63. Selective answering is a reasonable design choice for a small model.

Why dense still has a place

Most recent small-model gains have come from mixture-of-experts architectures, which carry large total parameter counts but activate only a fraction per token. MiniCPM5-2B goes the other direction: every one of its 2.6B parameters fires on every forward pass. The practical payoff is a smaller memory footprint rather than reduced compute per token, which is the tradeoff that matters on phones, laptops, and edge hardware where VRAM is the binding constraint. OpenBMB is iterating quickly on this efficiency-first line, and the full benchmark breakdown is on the Artificial Analysis model page. One version-comparison caveat: MiniCPM5-2B scored 23 on Intelligence Index v4.1.1 and 15 on v4.2, so cross-version numbers are not directly comparable.

When to reach for it

MiniCPM5-2B suits agents that call tools, fill out forms, or run multi-step knowledge workflows on constrained hardware. Workloads that demand frontier knowledge recall, competitive coding, or long-horizon terminal use will still be better served by larger models. Apache 2.0 licensing means no friction for commercial deployment.