NaiveAI published MIT-licensed weights, inference code, architecture and training details.
NaiveAI opens a 309B sparse model with native 1M context
2026-09-27Frontier watchTRIAL
NaiveAI released Naive-N0.5-Flash under MIT: a 309B MoE with 15.5B active parameters and native 1M-token context. The model replaces full-attention layers with a mix of sliding-window and sparse attention and publishes weights plus inference code. Deployment requires FP8-capable NVIDIA hardware and roughly 315 GB just for weights; benchmarks are author-reported.
Official · WeightsNaive-N0.5-Flash model card and weightsOfficial · ReferenceNaiveAI technical research
Availability: weights and inference code; no hosted inference provider
What this means for you
For long-context coding and research systems, the architecture is a concrete test of sparse attention without full-attention layers. The high hardware requirement and vendor-only evidence make it a targeted infrastructure trial, not a default model choice.
FRONTIER WATCH
Public artifacts and validated results are different.
The recorded fact is the public work and its artifacts. Author-reported outcomes do not establish independent reproduction. Next checks include usable code, task conditions and third-party validation.
Evidence & related material
Technical details & scores
| Architecture | 309B MoE, 15.5B active; 39 SWA and 9 DSA layers |
|---|---|
| Context | native 1M tokens |
| Training | 3.25T tokens after architecture conversion |
| License | MIT |
| Availability | weights and inference code; no hosted inference provider |
| Hardware | FP8-capable NVIDIA GPUs; approximately 315 GB for weights |
| Adoption | 1,497 monthly downloads and 145 likes checked 2026-10-03 |
Technical impact 4
Engineering value 3
Adoption signal 3
Maturity 2
Verification cost 5
Risk 3