MiniCPM5-2B: Install and Run OpenBMB's 2B Model That Outscores 4B Rivals

MiniCPM5-2B is OpenBMB's dense 2 billion parameter language model, released September 6, 2026 under Apache 2.0. On its own benchmark suite it averages 53.9 across reasoning, tool use and agentic tasks, beating every model tested up to 4B parameters, including Qwen3.5-4B's 51.1. It runs through vLLM, SGLang or Transformers, with GGUF, MLX and GPTQ builds for local hardware.

Written by Priya Raghunathan, local-AI hardware reviewer. I check parameter counts, file sizes and benchmark tables against the primary source before recommending an install path, because a model card's own comparison chart is still marketing until someone checks the numbers.

How MiniCPM5-2B compares to similarly sized models

MiniCPM5-2B is the second release in OpenBMB's MiniCPM5 line, following MiniCPM5-1B, and it is built for on-device assistants, coding agents and tool-use workflows rather than as a general chatbot. OpenBMB's own evaluation set compares it against similarly sized open models, LFM2.5-2.6B, Qwen3.5-2B and Gemma-4-E2B-it, plus a set of larger 4B-class models for reference. MiniCPM5-2B leads that comparison on average, and its biggest gaps show up in code, math, long context and agentic tasks rather than general knowledge.

"It remains competitive with 4B-class models overall, while showing its advantages over models of comparable size in coding, mathematics, long-context understanding, tool use, and agentic tasks."

Specs at a glance

How much hardware it actually needs

The model card does not publish a VRAM figure directly, but the GGUF repository on Hugging Face lists real file sizes for each quantization, and at 2.52 billion parameters none of them are large.

Install and run it

MiniCPM5-2B uses the standard LlamaForCausalLM architecture, so it loads directly in mainstream inference engines without a custom fork. OpenBMB's own quickstart covers three paths.

from transformers import AutoModelForCausalLM, AutoTokenizer
model_id = "openbmb/MiniCPM5-2B"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype="auto",
    device_map="auto",
)
messages = [{"role": "user", "content": "Who are you? Please briefly introduce yourself."}]
inputs = tokenizer.apply_chat_template(
    messages,
    tokenize=True,
    add_generation_prompt=True,
    enable_thinking=True,
    return_dict=True,
    return_tensors="pt",
).to(model.device)
outputs = model.generate(**inputs, max_new_tokens=128)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:], skip_special_tokens=True))

Run pip install -U "transformers>=5.6" accelerate torch before the script above. OpenBMB's recommended sampling settings are temperature 1.0 and top_p 0.95.

Tool calling for agents

SGLang is the backend OpenBMB recommends for tool and function calling. MiniCPM5-2B emits XML-style tool calls, and SGLang's built-in minicpm5 parser converts them to OpenAI-compatible tool_calls without extra glue code.

The tool-use and agentic scores in the table further up (66.6 on BFCL v4, 46.4 on SWE-bench Verified) are what OpenBMB points to as the reason to route function calling through SGLang rather than a generic parser.

GGUF, MLX and GPTQ if you don't want to run a server

For llama.cpp, Ollama or LM Studio, the quantized files live in the MiniCPM5-2B-GGUF repository, with sizes matching the table above. Apple Silicon users get a dedicated 4-bit MLX build, and there is a 4-bit GPTQ release for GPU setups that prefer that format. OpenBMB also publishes step-by-step cookbooks for each backend on its GitHub repo, which has passed 10,500 stars. If you just want to try it before installing anything, there is a demo Space on Hugging Face.

License and who it's for

Both the model weights and the GitHub repo are Apache 2.0, with no revenue cap and no attribution clause. OpenBMB's own disclaimer is worth repeating: the model has no autonomous intent, its outputs can be wrong or biased, and answers on topics like health, finance or law haven't been reviewed by experts. That makes it a reasonable pick for local assistants, coding agents and tool-use experiments on hardware too small for a 4B model, not a substitute for professional advice.

FAQ

How big is MiniCPM5-2B and what hardware does it need?

It has 2.52 billion parameters. The GGUF build ranges from 1.45 GB at Q4_K_M to 4.69 GB at full F16 precision, small enough to run without a discrete GPU on most current laptops and mini PCs.

Is MiniCPM5-2B actually better than Qwen3.5-4B?

On OpenBMB's own benchmark average, yes: 53.9 versus 51.1, despite Qwen3.5-4B having close to double the parameters. That comparison comes from OpenBMB's model card rather than an independent leaderboard, so it's a strong signal, not the final word.

Can I use MiniCPM5-2B for tool calling and agents?

Yes. OpenBMB recommends SGLang for this, since it ships a built-in parser for the model's XML-style tool calls. The card reports 66.6 on BFCL v4 and 46.4 on SWE-bench Verified, both ahead of the similarly sized models it's compared against.

What license is MiniCPM5-2B released under?

Apache 2.0, for both the Hugging Face model repository and the OpenBMB/MiniCPM GitHub repo. There is no revenue threshold and no required attribution.

Where to go from here

• New to quantization formats? Read the GGUF vs MLX vs NVFP4 explainer before picking a build.
• Comparing options at this size? See best GGUF models by VRAM tier for what else fits the same hardware.