Telperion releases 'Qwen3.8-27B-MLA' and 'GLA-g2' variants that cut KV cache to a quarter (community report)

Verified (fact)

Telperion released Qwen3.8-27B-MLA and Qwen3.8-27B-GLA-g2, variants of Qwen3.8-27B with Latent Attention applied. The quantized versions are calibrated for FP8 KV cache, and compressing the KV cache 4x reportedly allows nearly 4x more maximum context on consumer hardware and nearly 4x more concurrent throughput on servers.

Needs fact-check (claim)

Telperion said the MLA variant improves SWE performance by 6.58% (which it claims is statistically distinguishable) and uses up to 20% fewer reasoning tokens. Evaluation details are not in the materials.

Why it matters (interpretation)

If post-hoc conversion to shrink the KV cache of existing open models works, it could lower the cost of long contexts both locally and on servers.

Sources

• Community New model from Telperion: Qwen3.8-27B-MLA & Qwen3.8-27B-GLA-g2. 4x smaller KV cache, smarter, and more efficient reasoning. Reddit