Releases·Model Release
ByteShape Qwen 3.8 27B: To KL Diverge or Not to KL Diverge, Part 2: Metric Boogaloo
ByteShape released full ShapeLearn GGUF quantizations of Qwen 3.8 27B, reporting that its 3.84 bpw quant reaches 99.63% of BF16's aggregate score across eight benchmarks. The team also argues KL divergence is not a reliable proxy for quantization quality, in a paper accepted to the EMNLP 2026 Industry Track.
188 engagement1 sourceTue, Sep 15, 2026, 02:31 PM
AI Overview
- ByteShape released full ShapeLearn GGUFs for Qwen 3.8 27B; 3.84 bpw hits 99.63% of BF16's aggregate score across 8 benchmarks, 3.23 bpw hits 98.72%.
- All five ShapeLearn quants sit on the measured quality/speed-bpw frontier across six GPUs, beating Unsloth v3, ISTA-DASLab, AtomicChat and Bartowski comparisons; DFlash2 gave 1.34-2.10x and MTP 1.28-1.66x baseline throughput.
- Unsloth UD-IQ3_S showed ~20% lower KLD than a similar-size Lite model but scored 95.55% vs Lite's 97.33% of BF16, supporting ByteShape's KLD-fidelity paper accepted to EMNLP 2026 Industry Track.
Related
Social·Community
Qwen 3.8 27B IQ3-XXS on an RTX 3060: ByteShape 2.88BPW quant underperforms GSQ in one-shot coding tests
A user benchmarked two ultra-low-bit GGUF quants of Qwen 3.8 27B on an RTX 3060 12GB, comparing ISTA-DASLab's GSQ-RCO-IQ3-XXS against ByteShape's IQ3-XXS. Despite ByteShape's advertised high BF16 similarity, it failed tasks the GSQ quant completed in one shot. The result challenges ByteShape's KL-divergence-based quality claims for low-bit quantization.
A user benchmarked two ultra-low-bit GGUF quants of Qwen 3.8 27B on an RTX 3060 12GB, comparing ISTA-DASLab's GSQ-RCO-IQ3-XXS against ByteShape's IQ3-XXS. Despite ByteShape's advertised high BF16 similarity, it failed tasks the GSQ quant completed in one shot. The result challenges ByteShape's KL-divergence-based quality claims for low-bit quantization.
128 engagement1 sourceredditTue, Sep 22, 2026, 04:31 PM
Social·Community
I benchmarked 21 Qwen3.8 27B variants on 16GB VRAM
Reddit users are quantifying the quality of Qwen3.8 27B quantizations that fit in 16GB VRAM. One benchmarked 21 variants on an RTX 5080 against real C code, finding bartowski's IQ4_XS quant the best overall. A separate perplexity test found available NVFP4 GGUF quants underperform, all worse than Q4_K_XL at equivalent file sizes.
Reddit users are quantifying the quality of Qwen3.8 27B quantizations that fit in 16GB VRAM. One benchmarked 21 variants on an RTX 5080 against real C code, finding bartowski's IQ4_XS quant the best overall. A separate perplexity test found available NVFP4 GGUF quants underperform, all worse than Q4_K_XL at equivalent file sizes.
Updated Fri, Sep 4, 2026, 07:33 PM — A follow-up benchmark of 21 Qwen3.8 27B variants on real C code named bartowski's IQ4_XS the best overall quant.
356 engagement2 sourcesredditTue, Sep 1, 2026, 12:32 AM
Releases·Model Release
Telperion releases Qwen3.8-27B-MLA & GLA-g2 latent attention retrofits: 4x KV cache compression, +6.58% SWE-bench
Telperion announced two latent-attention retrofit variants of Qwen3.8-27B that compress the KV cache 4×, enabling near-4× larger context on consumer hardware or 4× higher server concurrency. The MLA variant posts a +6.58% gain on SWE performance with ~20% fewer reasoning tokens. Weights, quantized builds, and vLLM plugins are on Hugging Face and GitHub, with the method to be open-sourced.
Telperion announced two latent-attention retrofit variants of Qwen3.8-27B that compress the KV cache 4×, enabling near-4× larger context on consumer hardware or 4× higher server concurrency. The MLA variant posts a +6.58% gain on SWE performance with ~20% fewer reasoning tokens. Weights, quantized builds, and vLLM plugins are on Hugging Face and GitHub, with the method to be open-sourced.
28 engagement1 sourceredditWed, Oct 14, 2026, 04:47 PM
Social·Community
Prism-ML Bonsai 2 Qwen3.8 Quantization Hits 91.5% on Byteshape Composite Benchmark
Byteshape ran Prism-LM's newly released Bonsai 2 QAT models, based on Qwen3.8, through its own benchmark suite and added them to its Qwen3.8 quantization comparison. Bonsai 2 scored roughly 91.5% on the composite benchmark under Byteshape's methodology, with separate Instruct and Thinking evaluations. The goal is a common reference point, since vendors report numbers under differing runtimes, sampling settings, and suites.
Byteshape ran Prism-LM's newly released Bonsai 2 QAT models, based on Qwen3.8, through its own benchmark suite and added them to its Qwen3.8 quantization comparison. Bonsai 2 scored roughly 91.5% on the composite benchmark under Byteshape's methodology, with separate Instruct and Thinking evaluations. The goal is a common reference point, since vendors report numbers under differing runtimes, sampling settings, and suites.
204 engagement1 sourceredditFri, Sep 18, 2026, 04:01 PM
Releases·Tool Release
akopytko releases Blackwell-native NVFP4 quant of Qwen3.8 27B, 50% faster than Q4
A new GGUF quant of Qwen3.8 27B uses Blackwell-native NVFP4 with a prefill-optimized design, achieving 50% faster prefill than a Q4 quant of the same memory footprint and 4-7% faster than other NVFP4 quants on an RTX 5090. It includes a quantized MTP draft head and is available on Hugging Face.
A new GGUF quant of Qwen3.8 27B uses Blackwell-native NVFP4 with a prefill-optimized design, achieving 50% faster prefill than a Q4 quant of the same memory footprint and 4-7% faster than other NVFP4 quants on an RTX 5090. It includes a quantized MTP draft head and is available on Hugging Face.
166 engagement1 sourceredditFri, Aug 21, 2026, 09:19 AM