Naive-N0.5-Flash
#31 in Open-Source-SpraakmodelleNaiveai · flash · 2× · tolest 29. Sept. 2026
29
Momentum
Naive-N0.5-Flash is an Open-Weight language model published by NaiveAI (Beijing) with a Mixture-of-Experts architecture: 309 billion total parameters, of which 15.5 billion are active parameters per Token. The model exclusively uses a hybrid architecture of Sliding-Window-Attention (SWA) and lightweight DeepSeek Sparse Attention (DSA) in approximately a 5:1 ratio, without any Full-Attention layers, thereby achieving a native 1-million-Token context window. It is based on Xiaomi's open MiMo-V2.5 base model and was trained for coding and AI research/development tasks. Weights and Inference code are available under MIT license on Hugging Face; API access has been announced with tiered pricing.
Momentum-Verloop
02.07.30.09.
Features
| Key Benchmark (%) | SWE-bench Pro: 73.6% (3rd place behind Opus 5.5 at 89.9%) |
| Context Window (Tokens) | 1,000,000 tokens (native, without full attention) |
| License | MIT License (weights and inference code) |
| Multimodality | No multimodality specified; focused on text/code |
| Platform | Hugging Face (open-weight download), API announced (not yet live) |
| Price per 1M Tokens | $0.10 input / $0.40 output / $0.01 cache read (announced, API not yet live) |
| Release Date | September 27, 2026 |