Telperion Releases Qwen3.8-27B-MLA and GLA Variants with 4x KV Cache Compression
Telperion has released new latent attention variants of the Qwen3.8-27B model, namely Qwen3.8-27B-MLA and Qwen3.8-27B-GLA-g2. These models achieve a 4x reduction in KV cache size while improving software engineering benchmark performance and reducing reasoning token usage. This release significantly improves memory efficiency during text generation, allowing for a nearly fourfold increase in context length on consumer hardware or higher user concurrency on servers. It demonstrates that advanced attention mechanisms can be effectively applied to existing open-weights models without requiring expensive retraining. The MLA variant achieved a 42.1% score on SWE-bench Verified while using over 20% fewer reasoning tokens. Additionally, Telperion provided a grouped latent variant (GLA-g2) to address multi-GPU memory bottlenecks alongside custom vLLM plugins supporting various quantization levels down to INT4 and EXL3.
## BACKGROUND
KV cache (Key-Value cache) stores previously computed attention keys and values during text generation to avoid redundant calculations, but it consumes significant GPU memory as context lengths grow. Multi-head Latent Attention (MLA) addresses this bottleneck by compressing key-value tensors into a low-dimensional latent space, greatly reducing memory footprint while maintaining model performance.