MiniCPM5-2B Leads a Week of Small Model Releases: Benchmarks, Limits, and Edge Use Cases
OpenBMB's MiniCPM5-2B averaged 53.9 across its own benchmark set, ahead of larger 4B-class models. What the numbers mean, how Liquid AI's DSpark drafter speeds up small vision models, and why research on uncertainty features matters for deploying small models.
Late September 2026 brought a cluster of small language model releases, led by OpenBMB’s MiniCPM5-2B, a 2.5-billion-parameter model that its developers say outperforms every larger model in their comparison table. Around the same time, Liquid AI released a draft model that speeds up its small vision-language model, and a research paper showed how internal model features can flag answers that are likely to be wrong. Together they illustrate both the progress in small models and the questions teams should ask before deploying them on phones, laptops or edge hardware.
MiniCPM5-2B at a glance
According to the MiniCPM5-2B model card on Hugging Face, the model is the second in the MiniCPM5 series, after MiniCPM5-1B. It is a standard dense transformer (no mixture-of-experts routing) aimed at local assistants, coding agents, tool use and reasoning where a compact model is preferred.
| Spec | Value |
|---|---|
| Parameters | 2,516,756,480 (about 1.98 billion excluding embeddings) |
| Layers | 42 |
| Attention | GQA, 16 query and 2 key/value heads |
| Context length | 131,072 tokens |
| Licence | Apache 2.0 |
| Formats | BF16, GGUF (llama.cpp, Ollama, LM Studio), GPTQ 4-bit, MLX 4-bit, plus a DSpark draft model |
The benchmark claim, read carefully
OpenBMB compares MiniCPM5-2B with models of a similar size (LFM2.5-2.6B, Qwen3.5-2B and Gemma-4-E2B-it) and lists larger models for reference (Qwen3.5-4B, Granite-4.2-3B, Nemotron-3-Nano-4B, Gemma-4-E4B-it and LFM2.5-8B-A1B). Within that set, it reports an average of 53.9, ahead of all the larger models, the highest of which scored 51.1. Shattered’s analysis identifies that model as Qwen3.5-4B and notes the average spans 34 benchmarks across coding, math, instruction following, knowledge, long context, tool use and agentic tasks, with highlights including 69.1 on LiveCodeBench v6 and 86.5 on AIME 2026.
Three caveats apply:
- The scores are self-reported and the comparison set was chosen by OpenBMB. “State of the art” applies within that set, which the model card itself states.
- An average hides the spread. OpenBMB says its advantage is clearest in code, math, long context, tool use and agentic tasks; other categories may be closer or behind.
- A 2.8-point lead is modest, but it comes from a model with well under two-thirds the parameters of Qwen3.5-4B, which means lower memory use and cheaper hosting.
Shattered also lists other small releases logged that week, including Liquid AI’s drafter, Supersonic Labs’ Julia 1 decision model, Sarvam AI’s Saaras V4 speech-to-text model for 22 Indian languages, and Ornith AI’s 9B model.
Liquid AI’s DSpark: faster decoding for a small vision model
Liquid AI released an experimental DSpark draft model for its LFM2.5-VL-3B vision-language model. Speculative decoding uses a small “drafter” to propose several next tokens, which the main model then verifies in one pass. With matched sampling settings, the output is the same as the main model’s; only speed changes.
Key figures from Liquid AI:
- The drafter has about 280 million parameters, adding 8.9% to the deployed model.
- On an M5 Max MacBook Pro with MLX, decoding ran 2.30x to 3.13x faster depending on the task, and end-to-end latency improved 1.56x to 2.62x.
- On an M3 Ultra with llama.cpp, decoding improved 1.57x to 2.14x, end-to-end 1.30x to 1.77x.
- On a single H100 GPU with SGLang, decoding ran 2.04x to 2.66x faster.
Liquid AI is clear about the limits. Speculative decoding only speeds up the decode phase. For vision tasks, the image must first pass through a vision encoder and the model must process hundreds of visual tokens before generating anything, and on edge devices that stage can be a large share of total time. A big decode speedup therefore translates into a smaller overall gain. Higher sampling temperatures also lower the acceptance rate and therefore the speedup, and quantised models were outside the scope of the release.
OpenBMB publishes its own DSpark draft model for MiniCPM5-2B as well, so the same technique can be tested on the text model.
Knowing when a small model is wrong
A paper on sparse autoencoders, “Are LLM Uncertainty and Correctness Encoded by the Same Features?”, adds a useful angle. Working with Llama-3.1-8B and Gemma-2-9B, the authors found that features linked to a model’s uncertainty and features linked to actual correctness are largely distinct. Expressed confidence is therefore not a reliable signal of error.
Two results stand out:
- Suppressing “confounded” features that encode both uncertainty and error improved accuracy by 1.1% and reduced entropy by 75%, with effects carrying over between the ARC-Challenge and RACE benchmarks.
- Using just three of those features from one mid-network layer to predict correctness (AUROC about 0.79), the model could abstain on risky questions. That raised accuracy from 62% to 81% on the questions it answered, while answering 53% of them.
The paper did not test MiniCPM, and its models are larger. But the principle applies to any compact model: a deployment that can say “I am not sure” and hand off to a person or larger model may be more useful than one that always answers.
What this means for edge deployments
For teams considering a small model on devices or low-cost servers:
- Benchmark on your own tasks. Self-reported averages are a shortlist, not a decision. Test the two or three categories that matter for your product.
- Measure end-to-end latency, not just tokens per second. As Liquid AI shows, prefill and vision encoding can dominate on edge hardware.
- Try speculative decoding where drafters exist, but test at your real temperature and quantisation settings.
- Design an abstention path. Decide what the application does when the model is likely wrong, and measure how often that happens.
- Check the licence and formats. MiniCPM5-2B’s Apache 2.0 licence and GGUF and MLX builds make local testing straightforward.
For related coverage, see our look at diffusion-based language models with Mercury 2.5, the evolution of LLM context windows, and how open-weight models affect AI costs.
Our analysis
MiniCPM5-2B is a strong data point for the trend of small models closing the gap with larger ones, even allowing for the fact that its developers chose the comparison set. The more useful lessons for practitioners come from the surrounding releases: speed gains depend on which part of inference you accelerate, and reliability depends on knowing when a model should not answer. A small model that is fast, honest about its limits and tested on your own workload is worth more than a leaderboard average.