Qwen3.8-27B Runs on 12GB Memory, Unsloth Releases New Quantized Model

AI summary

Editorial policy

Unsloth released a quantized version of the 27B parameter Qwen3.8-27B model, enabling it to run on 12GB of memory.

Key points

•The model was released using a new quantization technique called Dynamic v3.1.
•It is distributed for free in GGUF format under the Apache License 2.0 on Hugging Face.
•The model can be run on environments such as the Unsloth Desktop app, llama.cpp, and Ollama.

Read the original

#LLM / Model Trends #Semiconductors & Infrastructure #Generative AI Tools & Use #Qwen3.8-27B#Unsloth#Ollama

#LLM / Model Trends#Semiconductors & Infrastructure#Generative AI Tools & Use#Qwen3.8-27B#Unsloth#Ollama

ITmedia AI+11/9/2026

Open model based on Qwen3.8-27B that "strengthens Western values" appears, also facing criticism for being "biased in the opposite direction" - From Hermes Agent developer

Nous Research released Arete-27B, based on Qwen3.8-27B and designed to strengthen Western values, facing criticism for opposite bias.

Hacker News11/6/2026

Offline survival assistant – local LLM and Kiwix, runs from a USB stick

A 100% offline survival assistant for Windows, Linux, and macOS that utilizes local AI via llama.cpp or Ollama and integrates Kiwix libraries.

Hacker News11/4/2026

Qwen 3.8 27B on my laptop can interpret model weights, too

Qwen 3.8 27B, running on a laptop, demonstrates the ability to interpret a research paper on LLMs natively reading and writing neural network weights.

Hacker News11/1/2026

Memdoor: A team of 5 Qwen3.8-27B models beats Fable 5 on LiveCodeBench Hard

Memdoor is a lightweight coding and automation agent in a single CLI binary that runs locally or over SSH, optimizes token usage, and automates durable CI and cron workflows.

Hacker News10/25/2026

Five Qwen3.8-27B models match Fable 5 Using GVS5H Harness

Using the GVS5H harness, five Qwen3.8-27B models achieve coding performance matching Claude Fable 5 on LiveCodeBench Hard, enabling Fable 5-level capabilities at one-fifth the cost or on a single GPU.