OpenBMB releases MiniCPM5-2B: a dense ~2.52B on-device model claiming 2B-class open-source SOTA

OpenBMB's MiniCPM5-2B, the second model in the MiniCPM5 series, is a dense 2.52B-parameter Apache 2.0 Transformer built for on-device deployment; it averages 53.9 across 34 benchmarks versus 51.1 for Qwen3.5-4B and trended as high as #4 on Hugging Face.

OpenBMB released MiniCPM5-2B, the second model in the MiniCPM5 series following MiniCPM5-1B, as a dense Transformer designed for on-device, local and resource-constrained deployment. The team claims 2B-class open-source state of the art and overall competitiveness with 4B-class models, highlighting coding, mathematics, long-context understanding, tool use and agentic capabilities. Per MarkTechPost, the model is a 2,516,756,480-parameter (~2.52B) dense LlamaForCausalLM with grouped-query attention and a 131,072-token context, released under Apache 2.0 and runnable via vLLM, SGLang, llama.cpp and Ollama, with GGUF weights also published on Hugging Face. It reportedly averages 53.9 across 34 benchmarks versus 51.1 for Qwen3.5-4B, with 97.1 on tau2-Bench Telecom (tool use), 69.1 on LiveCodeBench v6 and 46.4 on SWE-bench Verified (code), while being weaker on MMLU-Pro knowledge — so the 4B-competitiveness claim holds on average but not uniformly. Training reportedly combined 400B tokens of deep-thinking SFT, critic-based JustRL II RL teachers and on-policy distillation merging 16 RL experts, with datasets and intermediate checkpoints published alongside the weights. The release ships with a tech report (arXiv 2506.07900), a GitHub repository and an online demo. The GGUF variant trended #30 on Hugging Face on 2026-09-05 and the full weights trended #4 on 2026-09-06.

• Model: MiniCPM5-2B by OpenBMB, second in the MiniCPM5 series after MiniCPM5-1B (reports dated 2026-09-05 through 2026-09-07).
• Architecture: dense Transformer, 2,516,756,480 parameters (~2.52B despite the '2B' branding), LlamaForCausalLM with grouped-query attention and a 131,072-token context (MarkTechPost).
• License and runtime: Apache 2.0; runnable via vLLM, SGLang, llama.cpp and Ollama; GGUF weights available on Hugging Face for local/on-device use.
• Benchmarks: 53.9 average across 34 benchmarks vs 51.1 for Qwen3.5-4B; tau2-Bench Telecom 97.1 (tool use), LiveCodeBench v6 69.1 and SWE-bench Verified 46.4 (code); weaker on MMLU-Pro knowledge.
• Positioning: vendor-claimed 2B-class open-source SOTA and competitiveness with 4B-class models; the 34-benchmark average supports the Qwen3.5-4B comparison, but the MMLU-Pro weakness means it is not uniform across categories.
• Training recipe: 400B tokens of deep-thinking SFT, critic-based JustRL II RL teachers, and on-policy distillation merging 16 RL experts; open SFT/RL datasets and intermediate checkpoints released for verifiability.
• Artifacts: tech report (arXiv 2506.07900), GitHub repository and online demo accompany the release.
• Trending: GGUF variant at #30 on Hugging Face (2026-09-05); full weights at #4 (2026-09-06).