A site called Tiny League is tracking benchmark performance for small open-weight AI models, and the latest numbers reveal some unexpected results. The project, updated as of September 14, 2026, compiles scores across dozens of benchmarks for models released over the past several months, and the picture it paints challenges several common assumptions about model size and capability.
The scale of the benchmarking effort
Tiny League tracks over 70 distinct benchmarks, from established tests like GPQA Diamond, MMLU-Pro, and AIME to newer evaluations like Terminal-Bench, SWE-bench Pro, and MCP-Atlas. Each model in the database is scored across whichever tests its creators have published results for, and the site calculates an aggregate AA-Index score that attempts to normalize performance across the full set.
The models range from 2.5 billion parameters to 180 billion, covering both dense and mixture-of-experts architectures. Every model listed is open-weight, meaning its weights are publicly available for download and local deployment. This is the class of model that matters most for developers who need to run inference on their own hardware or in their own infrastructure without paying per-token API fees.
Dense models punch above their weight
The most striking finding is how well small dense models perform relative to their parameter counts. MiniCPM5-2B, a 2.5 billion parameter dense model released September 6, scores 86.5 on AIME 2026 and 97.1 on τ²-bench. Those numbers are competitive with models five to ten times its size. On LiveCodeBench, it reaches 69.1, and on HMMT February 2026 it hits 63.8, both respectable scores for a model that can run on a single consumer GPU.
Nanbeige4.2-3B, another small dense model at 4.2 billion parameters, scores 87.4 on GPQA Diamond and 82.8 on HMMT February 2026. Its AA-Index of 58.7 on the LCR benchmark and 52.2 on LiveCodeBench suggest it handles complex reasoning tasks far better than its size would predict. These are not toy models. They are serious general-purpose systems that happen to be small enough to run locally.
Mixture-of-experts models dominate throughput
The mixture-of-experts models tell a different story. Ling-3.0-flash, a 127.5 billion parameter MoE model released August 2, scores 93.2 on AIME 2026 and 87.7 on OmniDocBench 1.5. On BrowseComp, a web navigation benchmark, it reaches 82, and on BFCL v4, a function calling test, it scores 73. These are strong numbers across a wide range of task types.
DeepSeek-V4-Flash-DSpark, a 165.3 billion parameter MoE model released June 27, posts even more impressive results on certain benchmarks. It scores 3052 on Codeforces ELO, a competitive programming metric, and 1395 on GDPval-AA, suggesting strong performance on economic reasoning tasks. Its HLE score of 45.1 is the highest among all models in the database, and its Terminal-Bench 2.0 score of 56.9 indicates solid performance on systems-level tasks.
Vision models are catching up
The V badge in Tiny League marks models that support image input, and the vision-capable models are performing well. Qwen3.8-27BV, a 27.8 billion parameter dense model with vision support released August 5, scores 89.2 on GPQA Diamond and 90.3 on LiveCodeBench v6. Its AIME 2026 score of 94.6 is among the highest in the entire database.
Muse-Glimmer-30BV, a 29.8 billion parameter dense vision model released August 9, does not yet have benchmark results published across all tests, but its presence in the database signals that the open-weight vision model ecosystem is expanding rapidly. The pattern is clear: multimodal models are no longer a niche category. They are becoming the default.
The architecture split matters
The database makes it easy to compare dense versus mixture-of-experts architectures, and the differences are meaningful. Dense models tend to score higher on reasoning-intensive benchmarks like AIME and HMMT, where the full parameter set contributes to each answer. MoE models excel on throughput-oriented benchmarks and tasks that benefit from specialized subnetworks, like code generation and document understanding.
This is not a coincidence. MoE models activate only a fraction of their parameters for each token, which means they can be much larger in total while keeping inference cost proportional to the active parameters. The trade-off is that the sparse activation pattern can produce less consistent reasoning on tasks that require deep, sustained thought across many sequential steps. The benchmark data reflects this architectural reality.
License diversity and deployment implications
The license column reveals an important trend. Most of the top-performing models use Apache 2.0 or MIT licenses, which permit commercial use, modification, and redistribution without restrictions. This is significant for teams that need to deploy models in production without worrying about license compliance or usage limitations.
The exceptions are models from NVIDIA and Google, which use proprietary licenses. NVIDIA's Nemotron models and Google's Gemma models impose restrictions that may limit their use in certain commercial contexts. For teams that prioritize deployment freedom, the Apache 2.0 and MIT models represent the safer choice, and the benchmark data shows they are competitive on performance.
What this means for model selection
Tiny League provides a single place to compare open-weight models across a consistent set of benchmarks, which is valuable for teams making deployment decisions. The data suggests that the right model depends heavily on the use case. A team running inference on consumer hardware should look at the small dense models like MiniCPM5-2B or Nanbeige4.2-3B, which deliver strong reasoning performance at parameter counts that fit on a single GPU. A team running inference at scale on server hardware should consider the MoE models, which offer better throughput per active parameter.
The site also highlights how quickly the open-weight model landscape is changing. Models released just weeks apart can have dramatically different benchmark profiles, and the gap between the best and worst performing models in any given benchmark is large enough to matter in production. Teams that are not actively tracking these changes risk deploying models that are significantly outperformed by newer alternatives.
Tiny League is not a definitive ranking. No benchmark suite can capture the full range of capabilities that matter for real-world deployment. But as a snapshot of where open-weight models stand as of mid-September 2026, it is a useful starting point for anyone making decisions about which models to evaluate for their specific use case.