Is MiniCPM5-2B free to use commercially?

MiniCPM5-2B is published under Apache-2.0, so the weights can be used, modified and shipped in commercial products without a separate licence. OpenBMB also releases the RL training corpus behind it, UltraData-RL-2609, under Apache-2.0, though some upstream datasets inside it carry MIT, CC BY 4.0 and CC BY-SA 4.0 terms that still apply.

How does MiniCPM5-2B compare to Qwen3.5-4B?

On OpenBMB's own benchmark table MiniCPM5-2B averages 53.9 against 51.1 for Qwen3.5-4B, a model with roughly twice the parameters. The gaps are widest on long context, where MiniCPM5-2B scores 68.1 on NoLiMa versus 43.5, and on code reasoning, 69.1 on LiveCodeBench versus 56.4. Tool use is 97.1 versus 92.1.

Where is MiniCPM5-2B weak?

Artificial Analysis found MiniCPM5-2B weak on knowledge and coding tasks in its own testing, scoring 9% on both Humanity's Last Exam and Terminal-Bench. It rates the agentic side much higher, at an Elo of 831 on GDPval-AA v2. Artificial Analysis also calls it token-efficient for a reasoning model, using 19k output tokens per task.

Can MiniCPM5-2B run on a phone or a laptop?

Running MiniCPM5-2B locally is the point of the release. OpenBMB ships GGUF builds at 1.56 GB for 4-bit Q4_K_M, 2.68 GB for 8-bit Q8_0 and 5.04 GB at F16, which llama.cpp, Ollama and LM Studio can load. MLX builds cover Apple Silicon, and FlagOS adaptation extends the same weights to nine chip families.

How was MiniCPM5-2B post-trained?

OpenBMB describes the recipe for MiniCPM5-2B as RL plus OPD — reinforcement learning combined with on-policy distillation — and measures it at roughly 11 points of reasoning gain and 7 points on agentic tasks. The RL corpus is public: UltraData-RL-2609 holds 85,995 samples across maths, STEM knowledge, long context and code, each with a checkable answer or executable tests.

OpenBMB · 2026-09-07 · major

MiniCPM5-2B — a 2B open model that leads the sub-4B field

MiniCPM5-2B is OpenBMB's new 2.52B-parameter open model for phones and laptops. It averages 53.9 across the maker's benchmark set, ahead of Qwen3.5-4B at 51.1, and ships under Apache-2.0 with a 131K-token context.

GitHub card for the OpenBMB MiniCPM on-device language model repository

OpenBMB's new 2B model takes the top spot among open models under 4B parameters.

Quick facts

Maker	OpenBMB
Parameters	2.52B total, 1.98B non-embedding
Context window	131,072 tokens
License	Apache-2.0
Architecture	Dense LlamaForCausalLM, 42 layers, GQA
Released	September 7, 2026
Chip support	9 chip families via FlagOS

Benchmarks

Average across code, maths, long context and tool use

MiniCPM5-2B		53.9%
Qwen3.5-4B		51.1%
LFM2.5-2.6B		33.2%
Qwen3.5-2B		28%

source ↗

What is it?

MiniCPM5-2B is a 2.52-billion-parameter dense model built to run on phones, laptops and other hardware without a datacenter GPU. OpenBMB released it on September 7, 2026 under Apache-2.0, with a 131,072-token context window and 1.98B of its parameters outside the embeddings. Artificial Analysis ranks it highest among open-weight models under 4B parameters on its Intelligence Index v4.2.

How does it work?

The architecture is deliberately plain — a standard LlamaForCausalLM stack of 42 layers with 16 query heads and 2 key-value heads for grouped-query attention. Dense means every parameter is active on every token, so memory use stays predictable on a phone in a way a mixture-of-experts model does not. The post-training recipe is where MiniCPM5-2B differs: reinforcement learning paired with on-policy distillation, and OpenBMB published the RL corpus behind it as UltraData-RL-2609.

Why does it matter?

Local models usually give up tool use and long context to fit on a laptop. On OpenBMB's numbers that trade is much smaller here: 97.1 on τ²-Bench tool use and 68.1 on the NoLiMa long-context test, both ahead of Qwen3.5-4B at nearly twice the size. Quantized builds start at 1.56 GB, and FlagOS adaptation means the same weights run on Nvidia, Ascend, Kunlunxin and ARM-v9 hardware.

Who is it for?

people running local models on laptops and phones

Frequently asked questions

Is MiniCPM5-2B free to use commercially?: MiniCPM5-2B is published under Apache-2.0, so the weights can be used, modified and shipped in commercial products without a separate licence. OpenBMB also releases the RL training corpus behind it, UltraData-RL-2609, under Apache-2.0, though some upstream datasets inside it carry MIT, CC BY 4.0 and CC BY-SA 4.0 terms that still apply.
How does MiniCPM5-2B compare to Qwen3.5-4B?: On OpenBMB's own benchmark table MiniCPM5-2B averages 53.9 against 51.1 for Qwen3.5-4B, a model with roughly twice the parameters. The gaps are widest on long context, where MiniCPM5-2B scores 68.1 on NoLiMa versus 43.5, and on code reasoning, 69.1 on LiveCodeBench versus 56.4. Tool use is 97.1 versus 92.1.
Where is MiniCPM5-2B weak?: Artificial Analysis found MiniCPM5-2B weak on knowledge and coding tasks in its own testing, scoring 9% on both Humanity's Last Exam and Terminal-Bench. It rates the agentic side much higher, at an Elo of 831 on GDPval-AA v2. Artificial Analysis also calls it token-efficient for a reasoning model, using 19k output tokens per task.
Can MiniCPM5-2B run on a phone or a laptop?: Running MiniCPM5-2B locally is the point of the release. OpenBMB ships GGUF builds at 1.56 GB for 4-bit Q4_K_M, 2.68 GB for 8-bit Q8_0 and 5.04 GB at F16, which llama.cpp, Ollama and LM Studio can load. MLX builds cover Apple Silicon, and FlagOS adaptation extends the same weights to nine chip families.
How was MiniCPM5-2B post-trained?: OpenBMB describes the recipe for MiniCPM5-2B as RL plus OPD — reinforcement learning combined with on-policy distillation — and measures it at roughly 11 points of reasoning gain and 7 points on agentic tasks. The RL corpus is public: UltraData-RL-2609 holds 85,995 samples across maths, STEM knowledge, long context and code, each with a checkable answer or executable tests.

Try it

vllm serve openbmb/MiniCPM5-2B --port 8000