The AI optimization firm Unsloth released new GGUF files and quantized weights that enable Alibaba's Qwen3.8-27B model to operate on consumer hardware with minimal memory. The 1-bit quantized versions can run on systems with 8GB of RAM while retaining approximately 77% of the accuracy of the BF16 version. Additionally, the Unsloth Dynamic v3 GGUFs provide a 10% increase in top-1% accuracy, validated by a new metric called Divergence-300 which uses 300 unseen examples from DeepSWE and Terminal Bench to measure greedy accuracy across 32 tokens.
Tinker launched the Qwen3.8-27B model on its platform, providing access to native multimodal capabilities for images and video. The model is designed for professional work, research, and long-horizon agentic tasks, and features flexible thinking control alongside improved coding performance.
Key sources
- SOURCE@danielhanchen“Divergence-300 is a new metric which extends top-1% greedy acc to 32 tokens & more, and we used 300 unseen examples from Terminal Bench, DeepSWE and more”x.com
- SUPPORT@danielhanchen“We originally didn't want to release them but were shocked that they worked extremely well on our internal testing!”x.com
- SUPPORT@tinkerapi“It’s natively multimodal with images + video, has flexible thinking control, and is meaningfully better at coding, professional work, research, and long-horizon agentic tasks.”x.com
- SOURCE@alibaba_qwen“Strong enough to handle professional tasks. Small enough to run on your local machine.”x.com
- SUPPORT@alibaba_qwen“4 days to the top.🏆Thank you to every builder who pushed Qwen3.8-27B to #1 on Cline!”x.com