Open MoE LLM for coding trained from scratch on free compute. 3.45B total / ~0.5B active, Gated DeltaNet + full attention.
AI & ML interests
Open mixture-of-experts LLMs, code generation, long context, linear attention (Gated DeltaNet), training on free compute (JAX/TPU)
Recent Activity
View all activity
Organization Card
Polite Chainsaw
An independent AI project from Ukraine. We train Lewi Surge, an open mixture-of-experts language model for programming, from scratch on free and donated compute, one stage at a time.
- Lewi Surge Preview: 3.45B parameters, ~0.5B active per token, 128 experts (6 routed + 1 shared), Gated DeltaNet + full attention in a 3:1 ratio
- Every number is published as measured; checkpoints are released as they are, with their limitations stated
datasets 0
None public yet