Back to News

Qwen3.8 27B quantization benchmark: 4-bit holds, 1-bit collapses

284 points · 139 comments#quantization#benchmark#qwen#llm

A Hacker News post benchmarks quantizations of the Qwen3.8 27B model, reporting that 4-bit quantization maintains performance well, while 1-bit quantization leads to a collapse in quality. The post provides practical data for researchers and engineers considering model compression.

Coverage timeline

  1. Hacker Newsstared
    Benchmarking Qwen3.8 27B quantizations: 4-bit holds up, 1-bit collapsesHN

Related stories

  • ESP32S3 Cluster Runs 1.58-bit BitNet Language Model

    A Hacker News post demonstrates an ESP32S3 cluster running a 1.58-bit BitNet language model, showcasing ultra-low-bit inference on embedded hardware. The hands-on engineering demo highlights the feasibility of running quantized language models on resource-constrained devices.

    1 reports · 1 sources150 points · 31 comments

  • PrismML releases Bonsai 2 27B, compressing Qwen3.8 27B to 5.9 GB for smartphones

    PrismML has released Bonsai 2 27B, a compressed version of Alibaba's Qwen3.8 27B model. The model fits in 5.9 GB, making it suitable for smartphones, while retaining 98.2% of Qwen's benchmark scores.

    1 reports · 1 sources

  • Meta FAIR and UW propose byte-level distillation, predicting up to 4-point accuracy gain

    Researchers from Meta FAIR and the University of Washington propose a byte-level distillation method that converts a teacher model's token-level probability distribution into byte-level distributions by adding end-of-token markers. In distillation experiments using Llama 3-8B as the teacher, scaling-law predictions indicate that byte-level distillation will surpass token-level distillation as compute increases, with downstream average accuracy up to 4 percentage points higher.

    1 reports · 1 sources

  • Alibaba open-sources Qwen3.8-27B, a 27B multimodal dense model

    Alibaba's Qwen team released the open-weights Qwen3.8-27B, a 27B-parameter native multimodal dense model under the Apache 2.0 license. It supports 262K native context (extendable to 1M tokens) and outperforms Qwen3.7-Plus in coding and office tasks, with independent benchmarks showing it matching GPT-5.6 Luna on the Artificial Analysis Intelligence Index. The model surpassed 1 million downloads within days of release.

    11 reports · 4 sources1421 points · 788 comments

News updates every hour. Subscribe in your RSS reader to get new stories automatically.

Subscribe