Qwen 3.8 27B was released on August 14, 2026, with Shape Learn-Lite GGUFs published four days later. These Lite versions used smaller optimization budgets and fewer checks. Now the full Shape Learn models are complete and benchmarked alongside the Lite set.

The full Shape Learn models perform significantly better. GPU-5 reaches 99.63% of BF16's benchmark score, while GPU-4 achieves 98.72% at 11.0 GB versus 13.1 GB. All five models sit on the quality-speed frontier across six GPU comparisons.

Speculative decoding with MTP or DFlash2 boosts throughput across all models. DFlash2 is faster but requires more memory and lacks image input support. MTP works with image inputs and is better for VRAM-constrained scenarios.