Discussion on LatentMoE Sparsity and Scaling Efficiency

ivan_bezdomny · x · 2026-07-17

The thread focuses on sparsity designs like LatentMoE:

• The author has long argued that activation ratios can drop below 2%, with reasons to sustain scaling even at <1%.
• The cited text mentions that LatentMoE activates only 16/896 experts. Combined with Kimi Delta Attention and AttnRes, it reportedly achieves 2.5x more efficient scaling.
• It also notes that rumors of OpenAI and Anthropic moving towards significantly higher sparsity align with this trajectory.
• The poster adds a cost comparison: K3 has a per-token cost similar to V4, but being much larger, its overall cost is 17x higher, raising questions about profit margins.

Original post →

More from Models

• Rumor claims Kimi faked performance by serving Claude; DeepSeek new model surprises in evals — realsohamparekh · 2026-09-11
• GPT-5.6 writes well but is instantly forgettable, user complains — BasedRaddka · 2026-09-11
• Opus Refuses Protein Research Codebase Over 'Safety' Concerns, Dev Considers Rolling His Own — josephdviviano · 2026-09-11
• User Hails Unconfirmed 'DeepSeek 4.1 Flash' as an Inflection Point in LLMs — himanshustwts · 2026-09-11
• Terminal Bench v4: GLM-5.3 Leads at 41.9%, Kimi-K3 Underwhelms at 12.6% — Ok_Warning2146 · 2026-09-11
• GPT-6 Astra beats Factorio with enemies in 44 in-game hours at ~$4,500 API cost — liminal_bardo · 2026-09-11

Models channel →