MiniCPM5-2B

// Overview

MiniCPM5-2B is a small open-weight model from OpenBMB, released on 7 September 2026 under Apache-2.0 and sized to run on hardware people already own — a phone, a laptop, an edge box — rather than a datacenter GPU. The model card lists 2,516,756,480 total parameters, of which 1,981,982,720 sit outside the embeddings, and a 131,072-token context window.

The architecture is deliberately ordinary: a standard `LlamaForCausalLM` stack of 42 layers with 16 query heads and two key-value heads for grouped-query attention. Dense rather than mixture-of-experts means every parameter is active on every token, which keeps memory use predictable on a device with a fixed budget.

On OpenBMB's own comparison table it averages 53.9 across the published benchmark set, against 33.2 for LFM2.5-2.6B, 28.0 for Qwen3.5-2B and 24.6 for Gemma-4-E2B-it. Artificial Analysis, testing independently, places it at 15 on its Intelligence Index — "one point behind Ling 3.0 Tiny (16)" — with weaker knowledge and terminal scores (9% on both Humanity's Last Exam and Terminal-Bench v2.1) but an Elo of 831 on GDPval-AA v2, roughly 110 points ahead of Ling 3.0 Tiny.

OpenBMB ships the weights alongside GGUF, GPTQ and MLX builds and a DSpark speculative-decoding draft model, and lists vLLM, SGLang, Transformers, llama.cpp, Ollama, LM Studio, MLX and FlagOS multi-chip variants among supported runtimes.

// Benchmarks

This model's scores

1. AIME 202586.5%
2. MMLU-Pro70.8%
3. LiveCodeBench v669.1%
4. SWE-bench Verified46.4%

Scores on a 0–100 scale (25-point gridlines); higher is better. Each benchmark links to its published source.

// Strengths

• Runs on-device: a dense 2.5B model with quantized GGUF, GPTQ and MLX builds for llama.cpp, Ollama, LM Studio and Apple Silicon
• 131,072-token context, unusual at this size
• Leads its published comparison set on every transcribed row, including SWE-bench Verified at 46.4 against 6.0 or less for the peers listed
• Apache-2.0 weights, so commercial use needs no separate licence
• Broad runtime support — vLLM, SGLang, Transformers, llama.cpp, Ollama, LM Studio, MLX and FlagOS multi-chip variants

// Best for

• Local assistants on a laptop or phone where sending prompts to a hosted API is not an option
• Long-document work on-device, using the 131K context instead of chunking
• Edge and embedded deployments across non-Nvidia silicon through the FlagOS builds
• A draft model for speculative decoding, via the published DSpark variant