MiniCPM5-2B

ADDED 2026-10-17. Second model in the MiniCPM5 series, a dense LlamaForCausalLM (config model_type llama, architectures LlamaForCausalLM) post-trained with the card's RL + OPD recipe: a critic-based RL stage followed by on-policy distillation that merges 16 RL expert models, 5 of them agentic. The card says RL + OPD lifts reasoning and general benchmarks by an average of 10.96 points and agentic benchmarks by 6.96 over the pre-RL checkpoint. RUNNABLE: 2.52B dense, one BF16 checkpoint of about 5 GB, and the card ships GGUF, MLX 4-bit and GPTQ 4-bit builds, so it runs on any consumer GPU. PARAMS: the card states 2,516,756,480 parameters and 1,981,982,720 non-embedding parameters, and the HF API safetensors total matches the full count exactly, so params_b is 2.517 and active_params_b equals it (dense). flops_per_token_active = 2 x 2.516756480e9 = 5.034e9. GEOMETRY from config.json fetched 2026-10-17: hidden 2048, 42 layers, 16 query heads, 2 kv heads, head_dim 128 explicit (2048/16 = 128, consistent), so gqa_ratio 8. intermediate 6144, silu, vocab 130560, tie_word_embeddings false, bos 0, eos [1, 130073]. rope_theta 5e6 with rope_scaling null, so context_scaling_method is native. KV (bf16, every layer caches): 42 x 2 kv heads x 128 x 2 (K+V) x 2 bytes = 43,008 B/token (42 KiB). BENCHMARKS ARE VENDOR CLAIMS. The card compares against LFM2.5-2.6B, Qwen3.5-2B and Gemma-4-E2B-it in the 2B class and against larger models for reference; none of those rows are records in this catalog. BEATS llada2-2-mini, the only catalogued model in the sub-8B-active class with published numbers on these benchmarks: LiveCodeBench v6 69.1 vs 28.14, AIME 2026 86.5 vs 35.05, GPQA Diamond 70.2 vs 44.41, and MMLU-Pro 70.8, which llada2-2-mini does not report. Inside the card's own table it is also ahead of every listed 2B-class model on all four (Qwen3.5-2B scores LiveCodeBench v6 20.2, AIME 2026 29.0, GPQA Diamond 45.6, MMLU-Pro 64.3) and ahead of the card's Qwen3.5-4B row on LiveCodeBench v6 (69.1 vs 56.4) and AIME 2026 (86.5 vs 82.7), but behind it on GPQA Diamond (70.2 vs 77.1) and MMLU-Pro (70.8 vs 78.0). SWE-bench Verified is 46.4 and Terminal-Bench v2.1 is 8.6 in the same table, both vendor claims and both well below the larger models in that table. Column order on every row is MiniCPM5-2B, LFM2.5-2.6B, Qwen3.5-2B, Gemma-4-E2B-it, Qwen3.5-4B, granite-4.2-3B, Nemotron-3-Nano-4B, Gemma-4-E4B-it, LFM2.5-8B-A1B. Several cells carry a dagger the card uses for its own caveat; the numbers above are copied as printed.