Open Models · first seen 14 October 2026, 16:47 UTC
Telperion releases latent attention variants of Qwen3.8-27B
Cuts KV cache 4x on a 27B model, easing long-context inference costs.
Telperion released Qwen3.8-27B-MLA and Qwen3.8-27B-GLA-g2, latent attention variants that compress the KV cache by 4x for more efficient reasoning.
TelperionOpenOpen weights
- Name
- Qwen3.8-27B-MLA
- Params
- 27B