Qwen3.8-27B — Q8_0 Vision

High-quality Q8_0 quantization of Qwen3.8-27B packaged for Ollama with full vision support. Now updated with MTP support for faster generation. Clocked at 70 tps on RTX Pro 6000 Blackwell.

This build is intended for users who want higher fidelity than typical Q4/Q5 quantizations while still running Qwen3.8-27B locally.

Model Details

Run with Ollama

ollama run zhamm/Qwen3.8:27Bb-Q8-vision

Ollama will automatically download the model if it is not already installed locally.

About Qwen3.8-27B

Qwen3.8-27B is a dense multimodal model designed for strong general-purpose performance while remaining practical for local deployment.

Qwen3.8 improves on previous Qwen generations in several areas:

• Coding — code generation, debugging, analysis, and multi-step development tasks
• Professional Work — document analysis, technical writing, structured reasoning, and knowledge work
• Research — synthesis of complex information and long-context analysis
• Agentic Tasks — improved planning, tool use, and handling of feedback during multi-step tasks
• Vision Understanding — analysis of photographs, screenshots, diagrams, documents, charts, and other visual material
• Long-Context Work — native support for very large context windows
• Flexible Reasoning — supports both reasoning-intensive and faster non-thinking interactions

Why Q8_0?

The standard Ollama Qwen3.8-27B distribution uses a lower-bit quantization to reduce memory requirements.

This version uses Q8_0, requiring approximately 30 GB for the model package but retaining substantially more numerical precision than lower-bit quantizations.

Q8_0 is a good choice when model quality is more important than minimizing VRAM or system RAM usage.

Memory Requirements

The model weights themselves require approximately 30 GB.

Additional memory is required for:

• KV cache
• Vision processing
• Context
• Ollama runtime overhead
• Concurrent requests

A GPU with more than 30 GB of VRAM is therefore recommended for full GPU offload. Larger context windows can significantly increase memory consumption.

Systems with insufficient VRAM can still use Ollama’s CPU/GPU offloading, although inference performance will be lower.

Vision

This build includes the vision projector required for multimodal operation.

It can analyze:

• Photographs
• Screenshots
• Technical diagrams
• Charts and graphs
• Scanned documents
• User interfaces
• Equipment images
• Maps and plans
• Other visual content

Example tasks include asking the model to identify objects in an image, interpret a diagram, extract information from a screenshot, or reason about visual and textual information together.

Long Context

Qwen3.8-27B natively supports context lengths up to 262,144 tokens.

This Ollama build is configured with:

num_ctx 262144

Large context windows consume substantial additional memory. For systems with limited VRAM, reducing the context window may provide significantly better performance.

For example:

/set parameter num_ctx 32768

or:

/set parameter num_ctx 65536

may be more appropriate for general-purpose use.

Recommended Uses

This model is particularly well suited for:

• General-purpose AI assistant workloads
• Coding and software development
• Technical troubleshooting
• Document analysis
• Long-document summarization
• Research and synthesis
• Image and screenshot analysis
• Security and engineering applications
• Agentic workflows
• Tool-enabled applications
• RAG systems
• Complex reasoning tasks

Notes

Qwen3.8-27B is a dense 27B-class model, so its inference characteristics differ significantly from Qwen Mixture-of-Experts models where only a subset of parameters are active for each token.

The Q8_0 quantization prioritizes model fidelity over minimum memory consumption.

Performance will depend heavily on GPU architecture, available VRAM, context size, prompt length, and Ollama version.

Credits

Base model: Qwen3.8-27B by the Qwen Team

Qwen3.8 is released under the Apache 2.0 license.

This repository provides a Q8_0 Ollama-compatible build with vision support.