Qwen3.8-Flash-Next on Glyd

Run a 180-billion-parameter AI model on one 16 GB NVIDIA graphics card
Linux · 64 GB of RAM · free for personal use

_{A space shooter written in one prompt by the model on an RTX 4080 SUPER, unedited · the prompt, the file and how often it works}

This runs Qwen3.8-Flash-Next on a normal PC with Glyd's own engine. It's a large, smart AI model that usually needs a server. It chats, writes code and works with your apps and coding agents. Nothing leaves your PC.

How fast is it?

Measured on one PC: an RTX 4080 SUPER (16 GB), a Ryzen 9 7950X3D and 64 GB of RAM. A token is about ¾ of a word.

Run tools/bench.sh to measure your PC.

How smart is it?

Inspect AI, 150 GSM8K and 150 MMLU-Pro questions, thinking off:

The logs: results/v0.1.0.

What you need

Install

Let your AI set it up

Using an AI coding assistant (Claude Code, Cursor, Codex, ...)? Paste this into it:

Set up Qwen3.8-Flash-Next on Glyd on this PC for me: https://github.com/surya-koritala/glyd-qwen-3.8-flash - follow docs/AI_SETUP.md in that repository.

Or do it yourself

git clone https://github.com/surya-koritala/glyd-qwen-3.8-flash && cd glyd-qwen-3.8-flash
./setup.sh

It checks your graphics card, RAM and disk, installs the engine (no sudo, nothing else to install) and downloads the model (about 82 GB; if the download stops, run it again and it continues). It asks two questions, precise or fast and how much context, then opens the app in your browser at http://127.0.0.1:8080. Press Enter for the recommended answers.

Next time, run ./setup.sh again: it starts right away and downloads nothing twice. Press Ctrl+C to stop it.

Precise or fast?

• Precise (recommended, an 82 GB download) keeps more detail in the model, so its answers are closer to the original model's. It is the better choice for code: long programs come out right much more often.
• Fast (a 68 GB download) writes about 15-20% faster, with less accuracy: fine for chat, weaker for long code.

./setup.sh asks; to switch later, tools/download.sh --fast and tools/serve.sh --fast (or the same without --fast for precise).

Using it

• In the browser: http://127.0.0.1:8080 has Chat, a live Monitor of the model and your GPU, CPU and RAM, and About. The chat shows tables and code, draws flowcharts, and runs the web pages, games and animations the model writes in a live preview.
• Your apps and coding agents: add an "OpenAI-compatible" provider with the base URL http://127.0.0.1:8080/v1. Any API key and any model name work.
  • Apps that use Anthropic's API: http://127.0.0.1:8080/v1/messages (Claude Code: ANTHROPIC_BASE_URL=http://127.0.0.1:8080 claude).
  • AI tools can also start, stop and ask it through its MCP server: claude mcp add glyd -- ~/.glyd-qwen-3.8-flash/bin/glyd-engine qwenmcp --repo /path/to/glyd-qwen-3.8-flash.
• Thinking: off by default, which is the fastest. Turn it on in the chat's settings, or with reasoning_effort in your app.
• Long prompts: up to 262,144 tokens of context (./setup.sh asks). A 32,000-token prompt is read in about 7 seconds; follow-up messages reuse the conversation and start in under a second.
• One request at a time: others wait their turn.
• Is it fast on your PC? With it running, tools/bench.sh measures a chat and 4K, 16K and 32K prompts in about 5 minutes and writes bench-report.md to share.

Something went wrong?

• It stopped while downloading. Run ./setup.sh again. It continues where it stopped.
• "No NVIDIA driver found" or "580 or newer is needed". Install or update the NVIDIA driver (on Ubuntu: sudo apt install nvidia-driver-595), reboot, then run it again.
• It's very slow and the disk is busy. Your PC doesn't have enough free RAM. Close other programs (browsers use a lot).
• Port 8080 is already in use. It's probably already running. Or start it on another port: tools/serve.sh 8090.

Still stuck? Open an issue with what it printed. Found a security problem? Report it privately: SECURITY.md.

How does it work?

The model is a team of 512 small specialists ("experts") in each of its 48 layers, and each word needs only 10 of them. A 16 GB card can't hold them all, so Glyd shares the work across your PC:

• Your graphics card keeps the experts that are used most. Your RAM holds the rest, and your processor works on those at the same time.
• Compressed by use: Glyd stores the experts in 2 to 4 bits each, with more detail where the model needs it most.
• Guess, then check: the model's own small helper guesses the next few words, and the big model checks them all at once.
• Long prompts layer by layer: a 32,000-token prompt goes through the model one layer at a time.

License and credits

The model is Qwen3.8-Flash-Next by the Qwen team, under its own license; Glyd's compressed copy on Hugging Face carries the same license.

• This repository (the setup scripts, the chat app, the benchmarks and the documentation) is open source under the Apache License 2.0.
• The Glyd engine ships compiled in the releases, under the Business Source License 1.1: free for personal, research and other non-commercial use. Commercial use needs a license ([email protected]). Each version becomes Apache-2.0 four years after its release.

Contributions are welcome: CONTRIBUTING.md.