|

OpenBMB Releases MiniCPM5-2B: A 2.52B Dense Model Averaging 53.9 Across 34 Benchmarks and Built to Run On Device

ByRicardo

OpenBMB has launched MiniCPM5-2B, the second checkpoint within the MiniCPM5 sequence and the follow-up to MiniCPM5-1B. It is a dense causal language mannequin with 2,516,756,480 parameters, of which 1,981,982,720 sit exterior the embeddings. It makes use of 42 layers, grouped-query consideration with 16 question heads and 2 key/worth heads, and a local context window of 131,072 tokens. The structure is customary LlamaForCausalLM, so mainstream engines load it with no customized kernels and no model-code fork.

Is it deployable? Yes. The weights are Apache 2.0 and run by vLLM, SGLang, Transformers, llama.cpp, Ollama, LM Studio, MLX and FlagOS.

What the benchmark desk truly reveals

OpenBMB compares MiniCPM5-2B towards LFM2.5-2.6B, Qwen3.5-2B and Gemma-4-E2B-it in the identical dimension class, and lists Qwen3.5-4B, granite-4.2-3B, Nemotron-3-Nano-4B, Gemma-4-E4B-it and LFM2.5-8B-A1B for reference. Across 34 benchmark rows it averages 53.9. The greatest baseline in that set is Qwen3.5-4B at 51.1, then granite-4.2-3B at 42.7 and LFM2.5-2.6B at 33.2.

On code reasoning MiniCPM5-2B posts 69.1 on StayCodeBench v6 towards 56.4, and 46.4 on SWE-bench Verified towards 33.6. Tool use is the widest margin: 97.1 on τ²-Bench Telecom, 66.6 on BFCL v4, and 20.8 on τ³-Bench Banking towards 6.8. Long context is break up, with 68.1 on NoLiMa towards 43.5, however 59.0 on AA-LCR towards 61.0 and 43.7 on LongBench v2 towards 47.3. General information is the place the scale hole reveals: 70.8 on MMLU-Pro towards 78.0, and 8.9 on Humanity’s Last Exam towards 9.9. OpenBMB marks rows sourced from Artificial Analysis individually from internally reproduced ones.

Training recipe: SFT, then RL, then on-policy distillation

Training follows the UltraData tiered information administration methodology described in original research. Base coaching runs secure and decay phases, then mid-training adapts the mannequin to the goal information distribution. Post-training begins with 400B tokens of deep-thinking SFT, then trains specialised RL lecturers for math, code, agentic duties and writing utilizing the critic-based JustRL II algorithm.

The closing step is on-policy distillation. OPD merges 16 RL specialists, 5 of them agentic, right into a single shipped mannequin. At every response place it computes full-vocabulary reverse KL divergence between scholar and instructor logits because the benefit estimate, changing the verification-based benefit. It reuses the RL prompts as distillation information, so no new corpus is constructed. OpenBMB measures the RL plus OPD stage at 10.96 common factors on reasoning and common benchmarks and 6.96 factors on agentic ones.


The information is open too

Alongside the weights, OpenBMB launched Ultra-FineWeb, Ultra-AdvantageousWeb-L3, UltraX, UltraData-Code, UltraData-Math, UltraData-SFT-2605, UltraData-SFT-Agent-2609 with 500K agent samples, and UltraData-RL-2609 with greater than 80K RL samples. Intermediate checkpoints are revealed as nicely, overlaying Base, Midtrain and SFT-only, so the contribution of every stage will be measured immediately.

Summary

MiniCPM5-2B is a reputable on-device possibility for agentic and tool-calling workloads, not a common information mannequin. Its benefit is clearest on software use, coding brokers and NoLiMa-style long-context retrieval, and it trails bigger fashions on MMLU-Pro, GPQA-Diamond and MATH-500. The open information and intermediate checkpoints make the RL plus OPD declare checkable, which issues greater than the headline common.

Key Takeaways

  • 2.52B dense mannequin, 131,072 token context, Apache 2.0, customary Llama structure.
  • Averages 53.9 throughout 34 benchmarks, forward of Qwen3.5-4B at 51.1.
  • Strongest on software use, coding brokers and long-context retrieval; weakest on information.
  • Post-training pairs 400B SFT tokens with RL lecturers and on-policy distillation.
  • Pre-training, SFT and RL datasets ship alongside the weights.

Check out the HF, GitHub repo and Web. Also, be at liberty to comply with us on Twitter and don’t neglect to be part of our 150k+ML SubReddit and Subscribe to our Newsletter. Wait! are you on telegram? now you can join us on telegram as well.

Need to associate with us for selling your GitHub Repo OR Hugging Face Page OR Product Release OR Webinar and so on.? Connect with us

The put up OpenBMB Releases MiniCPM5-2B: A 2.52B Dense Model Averaging 53.9 Across 34 Benchmarks and Built to Run On Device appeared first on MarkTechPost.

Similar Posts