Unsloth Dynamic V3.1 Boosts Qwen3.8-27B Accuracy 5-10%
Unsloth has released quantized versions of Qwen3.8-27B, a 27-billion-parameter model with vision and reasoning capabilities, a 256K context window, and local execution on 17GB of RAM/VRAM. The model specializes in agentic coding, vision, and chat tasks.
On November 11, Unsloth deployed Dynamic V3.1, an updated quantization algorithm that improves accuracy by 5-10% at comparable sizes across benchmarks including Terminal Bench, DeepSWE, LiveCodeBench, GPQA Diamond, MMLU-Pro, and AIME25. The update applied to 2-bit through 6-bit quantizations; 3-bit variants performed equivalently and remained unchanged.
Unsloth also released NVFP4 quantization, which requires NVIDIA Blackwell GPUs and runs 1.5x faster than BF16 checkpoints while maintaining 92-97% accuracy—including 96.68% on code and 92.15% on chat. FP8 KV cache calibration provides 2x longer context lengths. NVFP4 deploys via vLLM; standard GGUF quants run on older GPUs through llama.cpp or Unsloth Desktop, which features automatic parameter tuning, tool integration, and code execution.
Read the full article
Continue reading at Hacker News →This is an AI-generated summary. Read the original for the full story.