Qwen3.8-27B runs locally on 17 GB RAM via Unsloth Dynamic GGUF

The new model in the Qwen family arrives on personal computers without compromising context size or vision abilities. The 27 billion-parameter version of Qwen3.8 is now available in GGUF format quantized by Unsloth, allowing execution on just 17 GB of RAM, a capacity already present in many laptops. Unsloth says it is the strongest model in its weight class, and NVFP4-quantized versions have also been added to the repository.

Under the hood sits an architecture that combines two attention mechanisms. The model comprises 64 layers with a hidden dimension of 5,120, organized into 16 repeating blocks. Each block contains three Gated DeltaNet layers (linear attention with 48 V heads and 16 QK heads, head dimension 128) followed by a classic Gated Attention layer (24 query heads, 4 key-value heads, head dimension 256). RoPE is applied in dimension 64, and the FFN has width 17,408. Training with MTP (multi-token prediction) adds inference speed without increasing model size.

The released files use Unsloth Dynamic V3.0 (preview), an early-access version of the company’s dynamic quantization engine. The developers claim it is SOTA performance for GGUF at present. The package includes support for the Developer Role, enabling the model to operate within agentic tools such as Codex; MTP for fast inference; execution and fine-tuning directly in Unsloth Desktop; and improvements to parsing of nested objects for more reliable tool calling.

Qwen3.8-27B is a native vision-language model: it understands images, STEM diagrams, documents and videos up to hours in length. Thinking mode is enabled by default but can be disabled per request; inference depth can be adjusted via the reasoning_effort parameter, and context from previous messages is retained through preserve_thinking. The original context length is 262,144 tokens, with a possible extension to one million tokens.

For thinking mode the developers recommend temperature=1.0, top_p=0.95, top_k=20, min_p=0.0, presence_penalty=0.0, repetition_penalty=1.0. For regular instruction mode (no thinking) they recommend temperature=0.7, top_p=0.80, top_k=20, min_p=0.0, presence_penalty=1.5, repetition_penalty=1.0. In environments that separate token ceilings for internal reasoning from final output, the recommendation is to allocate generous space for reasoning within the million-token window.

The combination of hybrid DeltaNet-Attention architecture, aggressive yet precise quantization, and built-in support for agentic tools places Qwen3.8-27B in a unique position: a model with encoding, research and long-term planning capabilities that runs on a laptop without cloud or expensive GPUs. For developers building local applications or researchers requiring full privacy, this represents a notable step toward independent execution of frontier-level models on consumer hardware.