OpenBMB model product
0 checkpoint via continual training (~2T tokens, 25% of training-from-scratch cost).
Updated Aug 12, 2026. Default version: MiniCPM-SALA
Technical details for the model's default version.
This profile uses the model's current scored version. Arena ratings and prices are shown separately.
Benchmark scores for MiniCPM-SALA.
| HumanEval | 95.1% | 01 | 66 | 100.0% | C | |
| MRCR 128K (2-needle) | 28.6% | 01 | 1 | 100.0% | C | |
| MRCR 128K (4-needle) | 19.6% | 01 | 1 | 100.0% | C | |
| MRCR 64K (2-needle) | 29.8% | 01 | 1 | 100.0% | C | |
| MRCR 64K (4-needle) | 20.6% | 01 | 1 | 100.0% | C | |
| MRCR 64K (8-needle) | 16.6% | 01 | 1 | 100.0% | C | |
| NoLiMa 128K | 23.9% | 01 | 1 | 100.0% | C | |
| NoLiMa 32K | 54.5% | 01 | 1 | 100.0% | C | |
| NoLiMa 64K | 43.0% | 01 | 1 | 100.0% | C | |
| RULER 1000K | 86.3% | 01 | 1 | 100.0% | C | |
| RULER 128k | 89.4% | 01 | 1 | 100.0% | C | |
| RULER 2048K | 81.6% | 01 | 1 | 100.0% | C | |
| RULER 512K | 87.1% | 01 | 1 | 100.0% | C | |
| RULER 64k | 92.7% | 01 | 1 | 100.0% | C | |
| MRCR 128K (8-needle) | 10.1% | 02 | 2 | 0.0% | C | |
| LiveCodeBench v5 | 60.5% | 04 | 9 | 62.5% | C | |
| MBPP | 89.1% | 04 | 33 | 90.6% | C | |
| CMMLU | 81.5% | 05 | 6 | 20.0% | C | |
| BBH | 81.5% | 08 | 12 | 36.4% | C | |
| AIME 2024 | 83.8% | 19 | 53 | 65.4% | C | |
| LiveCodeBench v6 | 52.0% | 44 | 53 | 17.3% | C | |
| IFEval | 76.3% | 57 | 65 | 12.5% | C | |
| AIME 2025 | 78.3% | 73 | 114 | 36.3% | C | |
| MMLU-Pro | 67.0% | 95 | 129 | 26.6% | C |
Preference and agent-evaluation results for the default version.
The default version does not have a matching Arena result yet.
Provider-specific output speed and catalog latency for MiniCPM-SALA. Runtime does not affect the capability score.
No provider-specific speed or latency record is linked to the default version yet.
Output Speed is generated output tokens received per second. Catalog latency is reported separately from observed provider TTFT.
Official vendor API pricing appears first, followed by individual provider offers.
The default version has no current input or output token prices.
Official prices use only the vendor's configured official Provider and positive standard USD PAYG rates. Third-party offers remain explicitly labeled.
Available versions of this model. The score column identifies the version used in the overall ranking.
| MiniCPM-SALA | 22.3 | 9.5B | N/A | N/A | No | Apache 2.0 |
Key information about MiniCPM SALA and its available data.
0 checkpoint via continual training (~2T tokens, 25% of training-from-scratch cost). 5x inference speed over dense baselines at 256K tokens.
With HyPE (Hybrid Positional Encoding) and NoPE in sparse layers, the model extrapolates to 2048K tokens despite a 520K training length, enabling 1M-token inference on consumer GPUs like the RTX 5090.
Data as of 2026-08-11.
Open a comparison with the three ranked models immediately above and below this model.
Recommendations prioritize the same model type and family, then the closest LLMBoard score.
Common questions about MiniCPM SALA.
MiniCPM SALA's default version was released on Feb 11, 2026.
No official standard PAYG price is currently available for MiniCPM SALA.
MiniCPM SALA was created by OpenBMB.
A context window is not available for the default version.
No. The default version is not marked as having publicly available weights.
No provider offering is currently linked to the default version.
Nearby ranked alternatives include Qwen3 VL 4B Thinking, Llama 3.3 Nemotron Super 49B, Qwen2.5 VL 72B.