OpenBMB has released MiniCPM5-2B, a 2.6 billion parameter dense reasoning model that scores 13 on the Artificial Analysis Intelligence Index. That score makes it the highest-performing open weights model with fewer than 4 billion total parameters, outperforming models such as Qwen3.5 9B, Granite 4.2 8B, Qwen3.5 4B, and gpt-oss-20b despite being significantly smaller.
The model is text-only (input and output), supports a 131,000-token context window, and is released under an Apache 2.0 license. It ranks first among 47 models in its size class on the Intelligence Index, placing ahead of every other open weights model with fewer than 4 billion total parameters.
What Makes MiniCPM5-2B Stand Out
At 2.6 billion parameters, MiniCPM5-2B is a dense model, meaning all of its parameters are active during every inference pass. This is distinct from Mixture-of-Experts (MoE) architectures, where only a subset of parameters fire for any given token. The practical implication is that MiniCPM5-2B’s memory footprint is its primary size advantage, while its compute requirements scale with the full 2.6B weight set.
What makes the result remarkable is the gap. MiniCPM5-2B scores 13 on the Intelligence Index while matching the performance of models four times its size. Qwen3.5 9B, which has roughly 9 billion parameters, also scores 13 on the same benchmark. Granite 4.2 8B, Qwen3.5 4B, and gpt-oss-20b all score lower despite having substantially more parameters to draw on.
The Artificial Analysis Intelligence Index v4.3 evaluates models across 10 benchmarks, including agentic knowledge work, real-world task automation, coding, terminal use, long-context reasoning, scientific coding, physics reasoning, professional document reasoning, and knowledge reliability. A score of 13 in that composite reflects a model that can hold its own across multiple evaluation dimensions.
Agentic Performance: The Model’s Standout Capability
MiniCPM5-2B’s greatest strength lies in its agentic capabilities, meaning its ability to interact with tools, APIs, and structured environments to complete real-world tasks. Across three agentic evaluations in the Intelligence Index, the model performs well above what its size would suggest.
AA-Briefcase
On AA-Briefcase, Artificial Analysis’s agentic knowledge work benchmark, MiniCPM5-2B achieves an Elo of 438. That places it second in its comparison set, behind only Ling 3.0 Tiny (485) and ahead of Granite 4.2 8B (324). This benchmark measures a model’s ability to carry out knowledge-intensive tasks in an agentic setting, and MiniCPM5-2B’s result puts it within striking distance of a model with three times its parameter count.
GDPval-AA v2
The GDPval-AA v2 benchmark tests models on real-world work tasks against a human baseline of 1,000. MiniCPM5-2B reaches an Elo of 831 on this evaluation, roughly 110 points ahead of Ling 3.0 Tiny (718) and approximately 180 points ahead of Granite 4.2 8B (647). For context, models at this scale typically score far lower: LFM2.5-2.6B sits at 204, and Gemma 4 E4B (Reasoning) at 178.
τ3-Banking
On the τ3-Banking agentic tool-use benchmark, MiniCPM5-2B ties with Ling 3.0 Tiny at 21%, sharing the top position. The next best model, Granite 4.2 8B, reaches only 8%. This benchmark measures how effectively a model can use tools in a banking-domain workflow, and MiniCPM5-2B’s joint-first result underscores its practical utility in structured, tool-dependent environments.
Outside the Intelligence Index, additional benchmarks reinforce this agentic strength. On the τ2-Bench Telecom benchmark, MiniCPM5-2B scores 97.1, and on BFCL v4 (Berkeley Function Calling Leaderboard), it reaches 66.6. These tool-use results position MiniCPM5-2B as one of the most capable models of its size for function-calling and API-interaction tasks.
Token Efficiency
For a reasoning model, MiniCPM5-2B is remarkably frugal with its output. It uses an average of 19,000 output tokens per Intelligence Index task, 11,000 of which are reasoning tokens. That ties it with Granite 4.2 3B (also 19k) as the most token-efficient model in the comparison set.
To put this in perspective, Ling 3.0 Tiny spends approximately 56,000 output tokens per task to gain one additional Intelligence Index point. Granite 4.2 8B uses 33,000. For deployment scenarios where output tokens directly affect cost and latency, such as on-device or edge inference, MiniCPM5-2B’s efficiency is a significant advantage.
Where the Model Falls Short
MiniCPM5-2B’s strengths are concentrated in agentic and tool-use tasks. In broader knowledge, coding, and long-context evaluations, the model shows clear limitations.
Knowledge and Reasoning
On MMLU-Pro, a widely used knowledge benchmark, MiniCPM5-2B scores 70.8%, compared to 78.0% for the best model in its size class. On Humanity’s Last Exam, a more challenging benchmark that tests deep reasoning and factual knowledge, it scores 9%. That places it seventh in its comparison set, behind Gemma 4 12B (Reasoning) at 16%. On AA-Omniscience, which measures knowledge reliability and hallucination rates, the model scores -12. This negative score is not primarily caused by inaccuracy; rather, the model attempts only 29% of questions, giving it a Non-Hallucination Rate of 78% and an accuracy of 8% on the questions it does answer. The -12 reflects a penalty structure that rewards answering correctly while penalizing both wrong answers and excessive abstention.
Coding and Terminal Use
On Terminal-Bench v4.0, which evaluates agentic coding and terminal use, MiniCPM5-2B scores approximately 9%, placing eighth in its set. On SciCode, a scientific coding benchmark, it reaches 26%, placing it second behind Granite 4.2 8B (31%). OpenBMB’s own evaluation across 34 benchmarks shows more encouraging results on code reasoning: 69.1 on LiveCodeBench v6 (against 56.4 for the next-best model in its class) and 46.4 on SWE-bench Verified (against 33.6). These results suggest that the model can handle code generation and reasoning reasonably well, but struggles with more complex, multi-step coding tasks where larger models have the advantage.
Document Reasoning
On GDP.pdf, Artificial Analysis’s professional document reasoning evaluation, MiniCPM5-2B passes only 1% of tasks outright, behind gpt-oss-20b (high) at 2%. On AA-LCR v1.1, a long-context reasoning benchmark, it scores 59%, placing fifth in the set, one point behind Ling 3.0 Tiny (60%). On NoLiMa, another long-context evaluation, it reaches 68.1 compared to 43.5 for the next-best model in its class, suggesting stronger performance on certain long-context tasks.
Comparing MiniCPM5-2B to the Competition
The chart and table below summarize how MiniCPM5-2B compares to models the user specifically highlighted as outperformed:
| Model | Total Parameters | Intelligence Index v4.3 |
|---|---|---|
| MiniCPM5-2B | 2.6B | 13 |
| Ling 3.0 Tiny | ~8B | 16 |
| Qwen3.5 9B (Reasoning) | 9B | ~13 |
| Granite 4.2 8B | 8B | ~11-13 |
| Qwen3.5 4B (Reasoning) | 4B | ~13-14 |
| Granite 4.2 3B | 3B | 11 |
| gpt-oss-20b | 20B | ~11-13 |
The pattern is clear: MiniCPM5-2B competes with and often surpasses models with two to eight times its parameter count, particularly in agentic evaluations. Where it loses ground is in knowledge-heavy and coding-heavy benchmarks, where larger models benefit from broader training data.
What This Means for On-Device AI
MiniCPM5-2B’s combination of a small parameter count, Apache 2.0 licensing, and strong agentic performance points toward a specific use case: running capable AI agents locally on consumer hardware. A 2.6B dense model can fit within the memory constraints of modern laptops, tablets, and even high-end smartphones, especially with quantization techniques.
The model’s token efficiency reinforces this direction. Running a reasoning model locally typically means paying for every token in terms of latency and compute. MiniCPM5-2B’s 19k output tokens per task, compared to 56k for Ling 3.0 Tiny, makes it a far more practical choice for edge deployment where response time matters.
For users and developers interested in small, efficient AI models, MiniCPM5-2B represents a meaningful step forward. Its weights are available on Hugging Face for self-hosting, it averages 53.9 across 34 benchmarks covering coding, math, long-context understanding, and tool use, and its Apache 2.0 license permits commercial use.
MiniCPM5-2B Technical Specifications
- Parameters: 2.6 billion (dense)
- Architecture: 42-layer Transformer with Grouped-Query Attention (16 query heads, 2 key-value heads)
- Context Window: 131,072 tokens
- Input Modality: Text only
- Output Modality: Text only
- Reasoning: Yes (extended thinking / chain-of-thought)
- Knowledge Cutoff: December 31, 2025
- License: Apache 2.0
- Released: September 7, 2026
- Creator: OpenBMB
The Bottom Line
MiniCPM5-2B demonstrates that raw parameter count is becoming an increasingly poor predictor of model capability. With 2.6 billion parameters, it matches or exceeds models with 4 to 20 billion parameters on a composite intelligence benchmark, and its agentic performance is particularly strong for its size class.
The model has clear weaknesses in knowledge, coding, and document reasoning that prevent it from being a drop-in replacement for larger models. But for agentic workflows, tool use, and on-device deployment, MiniCPM5-2B sets a new standard for what open weights models at this scale can achieve.
