Qwen 3.8 27B at 2x speed on a 5090

husky82 pts0 comments

GitHub - jtc268/balto-speedrunner: Qwen 3.8 27B at 2x the speed on one RTX 5090. Code at 150 tok/s. Chat at up to 300 tok/s. · GitHub

/" data-turbo-transient="true" />

Skip to content

Search/

Sign in<br>Sign upAppearance settings

You signed in with another tab or window. Reload to refresh your session.<br>You signed out in another tab or window. Reload to refresh your session.<br>You switched accounts on another tab or window. Reload to refresh your session.

Dismiss alert

{{ message }}

jtc268

balto-speedrunner

Public

Notifications<br>You must be signed in to change notification settings

Fork

Star

main

BranchesTags

Go to file

CodeOpen more actions menu

Folders and files<br>NameNameLast commit message<br>Last commit date<br>Latest commit

History<br>32 Commits<br>32 Commits

.github/workflows

.github/workflows

docs

docs

runtime

runtime

scripts

scripts

src-tauri

src-tauri

src

src

test

test

.gitignore

.gitignore

LICENSE

LICENSE

README.md

README.md

THIRD_PARTY_NOTICES.md

THIRD_PARTY_NOTICES.md

package-lock.json

package-lock.json

package.json

package.json

View all files

Repository files navigation

On a Mac? Get Balto for Mac

First launch downloads the inference engine and model. Keep at least 90 GB free.

How it works<br>Benchmarks<br>Support Balto

Balto turns Qwen 3.8 27B into a fast local coding agent for one RTX 5090. Expect roughly 150 tok/s during real coding runs and up to 300 tok/s on clean chat prompts.

Run the installer, approve Windows if it asks to enable WSL, and Balto handles the rest. It resumes setup after a required restart, preserves partial downloads, and opens the coding workspace as soon as the model is ready.

The Windows app uses Tauri and the system WebView2 runtime. Model inference runs in Docker, and Balto keeps its Node.js workspace runtime in its own app data directory.

Measured on our RTX 5090

Workload<br>Output speed<br>Context<br>Notes

Chat<br>Up to 300 tok/s<br>Short prompt<br>Warm model

Code<br>150 tok/s<br>Live agent session<br>Real tool calls

What the installer does

Confirms that the PC has one RTX 5090, enough free disk space, and current NVIDIA support.

Installs Docker Desktop in its official per-user WSL 2 mode, or starts an existing installation.

Downloads the lmsysorg/sglang:qwen38-27b image at the exact tested digest.

Creates a persistent Docker volume for model files, so interrupted downloads resume and updates do not erase weights.

Starts RadixArk/Qwen3.8-27B-NVFP4 with RadixArk/Qwen3.8-27B-DSpark using the tested 80K configuration.

Installs the coding workspace into Balto's private app directory.

Reminds the owner to unload other local models before starting Balto.

Applies future performance configuration updates without deleting the persistent model cache.

First launch requires a large download and at least 90 GB of free disk space.

Private remote steering

Balto can use Tailscale Serve after the owner signs in to Tailscale. It keeps the workspace bound to 127.0.0.1 and exposes private HTTPS endpoints only inside the user's tailnet. Balto does not enable Tailscale Funnel or open a public router port.

The onboarding screen shows the exact private URL and lets the owner turn remote access off without changing unrelated Tailscale routes.

Tested configuration

The inference arguments live in runtime/balto.ps1. The important settings are:

model Qwen 3.8 27B NVFP4<br>context length 80000<br>attention backend flashinfer<br>max running requests 1<br>speculation DSpark, FP8 draft<br>sampling temperature 0.6, top_p 0.95, top_k 20

Balto uses safe sampling defaults for coding. It does not force greedy temperature zero sampling, which can trap this model in repetitive reasoning loops.

Development

Requirements:

Windows 11

Node.js 22 or newer

Rust stable with the MSVC target

WebView2

npm install<br>npm run check<br>npm run dev

Build the NSIS installer:

npm run build

Release signing

Balto supports two different signatures:

Tauri updater signatures protect update artifacts and are required by the in-app updater.

Windows Authenticode identifies Adore LLC as the publisher and prevents the unsigned-app SmartScreen warning.

Every install shows its version in Settings. A green update arrow appears when GitHub publishes a newer signed release; one click verifies, installs, and relaunches it.

The release workflow uses Azure Artifact Signing when publisher credentials are configured. The in-app updater always verifies Tauri update signatures. Local development builds remain unsigned.

License and credits

Balto Speedrunner is proprietary software, copyright 2026 Adore LLC. All rights reserved.

The coding agent interface integrates MIT-licensed software from DeepSeek AI. Inference is powered by SGLang. Qwen model weights remain under their own license. See THIRD_PARTY_NOTICES.md for the full notices.

Balto Speedrunner is not affiliated with or endorsed by DeepSeek, Qwen, Alibaba, SGLang, LMSYS, NVIDIA, Docker, Tailscale, Microsoft, or OpenAI.

If Balto saves you setup time, buy me a...

balto model qwen coding github runtime

Related Articles