Telperion Releases Qwen3.8-27B-MLA and GLA-g2 Variants with 4x KV Cache Compression

Telperion has released latent attention variants of Qwen3.8-27B, specifically Qwen3.8-27B-MLA and Qwen3.8-27B-GLA-g2, which compress the KV cache by 4x. This compression enables a near 4x increase in maximum context on consumer hardware or a near 4x increase in concurrency on server hardware.

Performance and Efficiency

The Qwen3.8-27B-MLA variant demonstrates a statistically distinguishable +6.58% increase in SWE performance compared to the base model, while using upwards of 20% fewer reasoning tokens. The Qwen3.8-27B-GLA-g2 variant uses two groups to maintain 4x KV compression for dual-GPU users, addressing memory downsides associated with serving MLA on multi-GPU devices.

| Model | SWE-bench Verified (n > 1500) | Mean Output Tokens | KV Compression (TP=1, TP=2) | | :--- | :--- | :--- | :--- | | Qwen3.8-27B | 39.5% | — | — | | Qwen3.8-27B-MLA | 42.1% | 0.88x | 4.00x, 2.28x | | Qwen3.8-27B-GLA-g2 | 39.4% | 0.86x | 4.00x, 4.00x |

Quantization and Compatibility

Telperion provides post-training quantization recipes that exceed the fidelity of equivalently sized .gguf quants. The models are supported by an easy-to-install vLLM plugin, verified on RTX PRO 6000s, H100s, and 2x RTX 5060Ti 16GB.

| Quant | Memory (incl. MTP head/vision) | Top-1 Disagreement (lower is better) | | :--- | :--- | :--- | | BF16 | 55.7GB | — | | FP8 | 31.5GB | 0.06% | | INT4 | 22.9GB | 0.17% | | EXL3-4.0bwp* | 17.2GB | 0.46% |

*EXL3 requires a custom vLLM plugin.

Future Roadmap

The underlying method requires zero-retraining and is compatible with models using grouped query attention or similar variants. Telperion plans to open-source the method soon and intends to release an improved version 2 with a full benchmark suite, as well as sparse latent attention variants for Qwen4-27B upon its release.