What it is
MiniCPM5-2B is a small language model from the Chinese open-source group OpenBMB. Released on September 7, 2026, it follows MiniCPM5-1B and has 2.52 billion parameters in a standard Llama-style layout, so common tools such as llama.cpp, Ollama, LM Studio, MLX and vLLM load it without custom code. It reads and writes text, with a 131,072-token context.
OpenBMB trained it for local assistants, coding agents and tool use. Post-training used 400 billion tokens of step-by-step reasoning examples, followed by reinforcement learning and distillation. Alongside the final model it published the base, mid-training and fine-tuning checkpoints, GGUF, MLX and 4-bit GPTQ builds, a draft model for faster generation, and much of the training data, including 500,000 agent examples.
Everything is under Apache 2.0.
Our review
MiniCPM5-2B is the small model to beat for tool calling and short coding fixes, at least on its makers' numbers. OpenBMB's table averages 53.9 across 34 benchmarks against 51.1 for Qwen3.5-4B, a model nearly twice its size, with the widest leads on SWE-bench Verified (46.4 to 33.6), the BFCL v4 function-calling test (66.6 to 56.8) and the NoLiMa long-context test (68.1 to 43.5). OpenBMB ran most of these itself; only the GPQA Diamond and Terminal-Bench figures come from Artificial Analysis.
The weaknesses are knowledge and stamina. It trails the 4B model on MMLU-Pro (70.8 to 78.0) and GPQA Diamond (70.2 to 77.1). On Terminal-Bench 2.1, which measures long multi-step terminal work, it scores 8.6 against 25.8, and on SWE-bench Pro 14.4 against 28.2. That fits a model post-trained on 400 billion tokens of reasoning traces and 500,000 agent examples: good at the next tool call, poor at holding a plan across dozens of them.
For a model sold on running locally, the card is thin on the numbers that decide a deployment. It gives no memory, latency or tokens-per-second figure for any device, and no measured speed-up for its DSpark draft model. File sizes fill part of the gap: the official Q4_K_M GGUF is 1.56 GB and Q8_0 is 2.68 GB, which puts it within reach of phones and modest laptops by size, though nobody at OpenBMB has published a measured footprint.
Openness is what separates it from Gemma 4's small sizes, which sit at 24.6 and 31.2 on the same table. Base, mid-training and fine-tuning checkpoints are released under Apache 2.0 alongside much of the data, so a team tuning it to its own tool schema starts from a documented point. The card also warns that llama.cpp's default min_p of 0.05 can make it repeat itself, a detail that will catch out anyone who loads it with stock settings.
Where it earns its place
A mobile developer shipping an offline assistant inside an Android or iOS app
Turn typed or spoken requests into calls to the app's own functions without a server round trip
The 1.56 GB Q4_K_M GGUF or the LiteRT-LM build runs on the device, so there is no per-call cost and user data stays on the phone. Function calling is its strongest suit (66.6 on BFCL v4); keep tool chains short, because long multi-step runs are where it fails.
A platform engineer running an internal coding helper on one mid-range GPU
Triage and patch small, well-scoped bugs in CI before a person looks
vLLM 0.21 or SGLang with the minicpm5 tool parser serves the 5 GB BF16 weights. 46.4 on SWE-bench Verified is high for the size, but 14.4 on SWE-bench Pro argues for single-file fixes, with harder tickets routed to a larger model.
An academic ML group studying small-model post-training
Reproduce or ablate a current SFT, RL and distillation recipe on a modest budget
The released checkpoints for each stage, plus the UltraData SFT, agent and RL datasets, let a team fine-tune from any point and compare, under Apache 2.0 with no usage restrictions. Few labs publish this much at this size.
An operations manager at a field-service firm whose technicians work with patchy signal
Give engineers a laptop assistant that fills job sheets from rough notes
Running locally removes API spend and the data-protection review that sending customer notes to a cloud model would trigger. The trade is weaker general knowledge than 4B models, so anything beyond form-filling needs checking.
Before you commit
- Setup
- Loads as a standard Llama model in Transformers 5.6+, vLLM 0.21+, SGLang 0.5.16+, llama.cpp, LM Studio, Ollama and MLX; OpenBMB recommends temperature 1.0, top_p 0.95 and min_p 0.
- Integrations
- SGLang is the recommended backend for tool calling (--tool-call-parser minicpm5); official GGUF, 4-bit MLX, 4-bit GPTQ and LiteRT-LM builds, plus a DSpark draft model for speculative decoding.
- Data and privacy
- Apache 2.0 for code and weights, with no revenue cap or acceptable-use addendum. Self-hosted only, so prompts stay on your hardware; no Hugging Face inference provider hosts it.
- Cost at scale
- No hosted price exists, so cost is hardware: about 5 GB of BF16 weights, 2.68 GB at Q8_0 and 1.56 GB at Q4_K_M, which a single consumer GPU or a laptop holds with room for context. OpenBMB publishes no throughput figures to size a fleet.
- Limits
- 131,072-token context. Scores of 8.6 on Terminal-Bench 2.1 and 14.4 on SWE-bench Pro mark long agent runs as the weak point, and llama.cpp's default min_p of 0.05 can cause repetition.
What's good
- Apache 2.0, with base, mid-training and SFT checkpoints and data
- 66.6 on BFCL v4 and 46.4 on SWE-bench Verified (vendor-run)
- 1.56 GB at Q4_K_M; loads in every common local runtime
- 131K-token context
What's not
- Trails Qwen3.5-4B by about 7 points on MMLU-Pro and GPQA Diamond
- 8.6 on Terminal-Bench 2.1: weak on long agent runs
- No published memory, latency or throughput figures
- Repeats itself under llama.cpp's default sampling
Best for
Developers putting tool-calling or coding helpers on phones, laptops and edge hardware, and researchers who want fully documented small-model training.
Alternatives
MiniCPM5-2B: common questions
- Is MiniCPM5-2B worth it?
- A 2B model that calls tools and patches code like a 4B one, and knows noticeably less than one. We rate it 4.0 out of 5.
- How much does MiniCPM5-2B cost?
- Free weights and training data under Apache 2.0.
- What are the downsides of MiniCPM5-2B?
- Trails Qwen3.5-4B by about 7 points on MMLU-Pro and GPQA Diamond; 8.6 on Terminal-Bench 2.1: weak on long agent runs; no published memory, latency or throughput figures; repeats itself under llama.cpp's default sampling.
- Who is MiniCPM5-2B best for?
- Developers putting tool-calling or coding helpers on phones, laptops and edge hardware, and researchers who want fully documented small-model training.
- How much does MiniCPM5-2B cost for a team or in production?
- No hosted price exists, so cost is hardware: about 5 GB of BF16 weights, 2.68 GB at Q8_0 and 1.56 GB at Q4_K_M, which a single consumer GPU or a laptop holds with room for context. OpenBMB publishes no throughput figures to size a fleet.
- Is MiniCPM5-2B safe for company data?
- Apache 2.0 for code and weights, with no revenue cap or acceptable-use addendum. Self-hosted only, so prompts stay on your hardware; no Hugging Face inference provider hosts it.
- What are the best MiniCPM5-2B alternatives?
- The closest alternatives we have reviewed are SmolLM3-3B, Granite 4.2 and Gemma 4.
Answers checked Oct 22, 2026.
Make MiniCPM5-2B? Show this rating on your site
