Alzeo
GitHub

Alzeo / Models / Qwen3.8 27B

Run Qwen3.8 27B on a DGX Spark

Text and code, a 1M-token window, by Qwen. NVFP4 (RadixArk) with a DFlash2 drafter, on SGLang 0.5.22.

Qwen's 27B in NVFP4 (RadixArk's build), served by SGLang 0.5.22 with a DFlash2 drafter, with a window of a million tokens. On an ASUS Ascent GX10 with 128 GB: 65.2 to 68.5 tok/s, greedy median in two runs, one request (2026-11-05).

Measured

65.2 to 68.5tok/s

greedy median in two runs, one request (2026-11-05)

  • 44.0 tok/s alone and 80.6 in all at 4 requests on 8,192-token prompts
  • A window of a million tokens
  • GSM8K 99.2% on 250 problems
Measured on
an ASUS Ascent GX10 with 128 GB, the project's reference box
How
./bench.sh and SGLang's official benchmark, 2026-11-05 (v1.44.0)
Source
BENCHMARKS.md

Load it

From the cockpit: open the Lanes view, pick Qwen3.8 27B and press Load. From a terminal, in the repo folder on the box:

./switch-model.sh stock   # switches the boot, then prints the commands to run

switch-model.sh never restarts an engine itself: it prints the exact stop and start commands. The cockpit's Load runs them for you.

Other checkpoints of the same lane load the same way: fp8, uncensored, uncensored-fp8.

On a 64 GB DGX Spark

Yes, without the 1M window: two requests at a time, 59,625 tokens of KV (one prompt of up to 54,855 tokens), 67.2 tok/s greedy median, measured at a 64 GB budget on the reference box. What runs on 64 GB

Use it from your tools

The box serves OpenAI- and Anthropic-compatible APIs on port 30001, with an API key. Claude Code, opencode, Open WebUI and the OpenAI and Anthropic SDKs point there instead of a cloud: see Claude Code with a local model and opencode with a local model.

Install it

One command, on stock DGX OS. The same command updates it later.

curl -fsSL https://get.alzeo.ai | bash

Questions about Qwen3.8 27B on a DGX Spark

Can a DGX Spark run Qwen3.8 27B?

Yes. Qwen's 27B in NVFP4 (RadixArk's build), served by SGLang 0.5.22 with a DFlash2 drafter, with a window of a million tokens. The installer of Alzeo sets it up, and the cockpit loads it with one button.

How fast is Qwen3.8 27B on a DGX Spark?

65.2 to 68.5 tok/s, greedy median in two runs, one request (2026-11-05), measured on an ASUS Ascent GX10 with 128 GB. 44.0 tok/s alone and 80.6 in all at 4 requests on 8,192-token prompts. A window of a million tokens. GSM8K 99.2% on 250 problems.

Does Qwen3.8 27B run on a 64 GB DGX Spark?

Yes, without the 1M window: two requests at a time, 59,625 tokens of KV (one prompt of up to 54,855 tokens), 67.2 tok/s greedy median, measured at a 64 GB budget on the reference box.

The other lanes