ByteShape has released its full ShapeLearn set of quantized Qwen 3.8 27B models, expanding an earlier Lite release with five GGUF variants designed for different balances of accuracy, size and inference speed. The company says all five models landed on the measured quality-speed frontier in its tests, meaning no other configuration plotted was both faster and more accurate.
The release targets users who want to run the dense 27-billion-parameter model on local GPUs. ByteShape recommends its GPU-5 variant where memory permits, reporting an aggregate benchmark result equal to 99.63% of the BF16 reference. Its smaller GPU-4 model remains close to 99% while leaving more room for context and runtime buffers. The company cautions that a model fitting into a card's nominal memory does not guarantee every context length or serving setup will fit.
Testing covered six Nvidia GPUs. On an RTX Pro 6000, ByteShape measured GPU-5 at 90.4 tokens per second. The same variant reached 93.7 tokens per second on an RTX 5090 and 45.8 on an RTX 4090. GPU-4 was faster on the 4090 at 49.5 tokens per second, while giving up less than one percentage point in the reported aggregate score. On tighter-memory RTX 4080 and RTX 5060 Ti systems, GPU-4 reached 52.4 and 33.1 tokens per second respectively; GPU-5 reached 45.7 and 29.1.
Those figures come from ByteShape's own evaluation and should be read as vendor benchmarks rather than independent validation. The comparison included models from AtomicChat, ISTA-DASLab, Prism-ML, Unsloth and others, though ByteShape noted that newer Bartowski releases arrived after testing and were excluded. Prism-ML's smallest Ternary Bonsai 2 variants were the fastest points in the charts, but had lower aggregate scores and required a custom runtime build.
ByteShape also tested speculative decoding. Every GGUF bundles an MTP draft head, while DFlash2 uses a separate 1.1 GB draft model and requires llama.cpp build 10658 or later. Across the lineup, the company measured DFlash2 at 1.34 to 2.10 times baseline next-token throughput and MTP at 1.28 to 1.66 times baseline.
The full release follows ShapeLearn-Lite, which ByteShape produced shortly after Qwen 3.8 27B appeared in August using a smaller optimization budget. The new results also underline a distinction in model compression: lower distribution divergence from a BF16 reference does not necessarily translate into better task performance. ByteShape says its rankings combine throughput with aggregate benchmark results, giving local-model users a more practical basis for choosing among file size, available context and response speed.
