Optimizations and New Quantizations for Qwen 3.8 27B
The open-source large language model ecosystem has seen a series of technical developments surrounding the Qwen 3.8 27B model, focusing on quantization efficiency and reasoning token reduction. Recent releases from ByteShape and UkisAI have introduced new methods for deploying the model on consumer-grade hardware, though performance trade-offs regarding accuracy and “hallucination” rates remain a subject of debate among users.
ByteShape Releases ShapeLearn GGUF Quantizations
ByteShape has announced the release of its full ShapeLearn GGUF series for Qwen 3.8 27B. According to company documentation, the new quantizations aim to sit on the quality-to-speed frontier across various GPU architectures, including NVIDIA RTX 3090, 4090, and 5090 series.
The company reported that its 3.84 bpw (GPU-5) quantization reaches 99.63% of the BF16 aggregate score across eight benchmarks, while the 3.23 bpw (GPU-4) version reaches 98.72%. ByteShape noted that their DFlash2 implementation delivered between 1.34× and 2.10× baseline throughput, while Multi-Token Prediction (MTP) provided a 1.28× to 1.66× increase in throughput when using temperature sampling.
In a technical discussion on Reddit, ByteShape addressed the relationship between Kullback–Leibler divergence (KLD) and model performance. The company noted that while lower KLD can prevent a quantization from “falling over the edge,” it is not a definitive indicator of task performance. ByteShape has submitted a paper on KLD and quantization fidelity to the EMNLP 2026 Industry Track to further explore these metrics.
UkisAI Releases Swift-Qwen3.8-27B
A separate development involves the release of Swift-Qwen3.8-27B by UkisAI. This model is designed to address the high latency associated with the “thinking” or reasoning tokens produced by the base Qwen 3.8 model. By utilizing reinforcement learning to penalize reasoning-marker tokens, the model aims to reduce the number of tokens required to reach a conclusion.
Unverified benchmarks shared by the developer claim that the model uses approximately 30% to 50% fewer reasoning tokens than the base model while maintaining comparable performance. In a comparative evaluation posted to Reddit, one user reported that the Swift version achieved a Pass1 rate of 30.8% compared to the base model’s 27.1%, while reducing total completion tokens from 12,547 to 7,301.
Performance Trade-offs and Technical Observations
Independent testing of the Swift-Qwen3.8-27B model has highlighted potential drawbacks to reduced reasoning. In a detailed technical report, a user testing the model on dual RTX 3090s found that while the model successfully reduced the median number of thinking characters, it exhibited a higher tendency to “bluff” or guess when faced with factual questions.
The report noted that in a test of 2,400 factual questions, the Swift model had a higher hallucination rate (48.0%) compared to the base model (44.3%). The tester observed that the Swift model tended to provide incorrect answers in instances where the base model would have opted to abstain from answering. However, the report concluded that instruction-following capabilities remained stable, with the Swift model scoring 66.1% on IFBench compared to the base model’s 64.0%.
Additionally, technical analysis of the Swift model’s Multi-Token Prediction (MTP) head suggested that using the embedded Q8_0 MTP head may result in lower tokens-per-second compared to using a standard external Q4_0 draft model, due to the higher computational cost of the larger draft weights.