Windows tutoria has been published

Full tutorial link:https://youtu.be/X8Tr2j5N07c

Run Qwen3.8-Flash-Next (125B) Locally + Unrestricted (Abliterated) Version Setup

1. Project Overview

Strata is an open-source inference engine specifically optimized for the Qwen3.8-Flash-Next model. It allows you to run this 125-billion-parameter model on consumer-grade GPUs (starting from around 8GB VRAM) by intelligently distributing the model across VRAM, system RAM, and SSD.

Key Features:

Supports both NVIDIA and AMD graphics cards
Typical generation speed: 50–100+ tokens/s (depending on GPU)
Image input support, long context, OpenAI & Anthropic compatible API
Can be paired with a control vector for near-uncensored / abliterated behavior

2. Hardware Requirements

Component	Recommendation
VRAM	8GB minimum (12GB+ recommended for smoother performance)
System RAM	48GB minimum, 64GB recommended
Storage	SSD recommended, 80GB+ free space
OS	Windows 10/11 or Linux

Memory Estimation:
Shard 1 (expert weights) is loaded into system RAM, plus roughly 10GB for OS overhead.

48GB RAM → Suitable for Q2_0 / IQ2_XS
64GB RAM → Can run IQ3_XXS / IQ3_S

3. Manual Installation of Strata (Base Version)

Step 1: Download the Project

Download the latest source code from GitHub:

https://github.com/Niko1221/Strata

Extract it to any folder, for example: D:\AI\Strata

Step 2: Run the Installer

Double-click START-HERE.bat and follow the prompts:

First run will automatically install a user-level Python environment (no admin rights required)
Model version: Recommended to choose Swift 1.5 (shorter thinking, faster responses)
Quantization size:
- 48GB RAM → Choose IQ2_XS
- 64GB RAM → Recommended IQ3_XXS or IQ3_S
Set context length and vision options according to your needs

After installation, your browser will automatically open the local chat interface (default: http://127.0.0.1:8080).

4. Quantization Options Reference

Quantization	Approx. RAM Requirement	Speed	Quality	Recommended Use
Q2_0	37.6 GB	Fastest	Good	Maximum speed
IQ2_XS	39.2 GB	Fast	Better	Daily recommended
IQ3_XXS	47.0 GB	Medium	Excellent	Quality priority
IQ3_S	54.8 GB	Slower	Best	Closest to original model

5. Manual Setup of Unrestricted Version (Abliterated)

Using the community-made Control Vector method, you can achieve near-uncensored behavior on your already-downloaded official model without re-downloading tens of gigabytes of files.

Steps:

Download the Control Vector file (~480 KB)
Repository: alesha-pro/Qwen3.8-Flash-Next-abliterated-GSQ-RCO-Strata-GGUF
Download: Huihui-Qwen3.8-Flash-Next-refusal-direction-r.gguf
Place the file
Put the downloaded file in the same folder as your model GGUF shards.
Example: Strata\models\swift-IQ3_XXS\

Edit the configuration file
Open the corresponding config file (e.g. strata-swift-iq3_xxs.json) and add the following parameters at the end of the "args" array:

"--control-vector-scaled",
"YOUR_FULL_PATH\\Huihui-Qwen3.8-Flash-Next-refusal-direction-r.gguf:1.0",
"--control-vector-layer-range",
"4",
"44",
"--cvec-mode",
"project",
"--cvec-dir",
"per-layer"

Save and restart
Save the JSON file and restart Strata for the changes to take effect.

Tip: The default strength of 1.0 works well for most users. To temporarily disable the effect, change :1.0 to :0 or remove the related lines.

6. Frequently Asked Questions

Can it run on 8GB VRAM?
Yes, but sufficient system RAM is important. Performance will be limited by memory bandwidth.
Does it support AMD GPUs?
Yes, though NVIDIA currently has better optimization.
How to improve speed?
Use a GPU with more VRAM, choose a lower quantization (e.g. IQ2_XS), and enable the Swift 1.5 version.
Control vector not working?
Check that the file path is correct, the JSON is valid, and that you have fully restarted Strata.

7. Useful Links

Strata Official Repository: https://github.com/Niko1221/Strata
Control Vector Repository: alesha-pro/Qwen3.8-Flash-Next-abliterated-GSQ-RCO-Strata-GGUF
Performance Details: See docs/DETAILS.md in the GitHub repository

This is the companion text for the video tutorial. It is recommended to follow along with the video.

Case Study : 3D Skiing Game

Create a complete, self-contained single HTML file that implements a 3D night skiing game using Three.js loaded via CDN.Requirements:Technical Constraints

Output only one complete HTML file (no external files except Three.js CDN).

Load Three.js from a reliable CDN (e.g. https://cdnjs.cloudflare.com/ajax/libs/three.js/r128/three.min.js).

Use Web Audio API to generate wind sound procedurally (no external audio files).

Fullscreen canvas with no page margins or white borders.

Clean, modern code with proper game loop and state management.

Visual Scene

Dark night sky background with animated stars and a visible moon.

Infinite downward-sloping snow track that continuously generates new segments.

Street lamps positioned on both sides of the slope, casting light onto the snow.

Atmospheric fog that increases with distance.

Randomly placed obstacles on the track: pine trees, rocks, and ski flag gates (two poles with a crossbar).

Distant mountain silhouettes on the horizon.

Third-person camera positioned behind and slightly above the skier. Camera should gently tilt when the skier turns.

Player Character

Construct the skier from simple geometric primitives:

Body, head, two ski poles, and two skis.

Skier body should lean/tilt when turning left or right.

Leave visible snow tracks behind the skis.

Emit particle snow spray from the skis when moving at high speed.

Gameplay Mechanics

Controls:

Left/Right Arrow: Steer left/right

Up Arrow: Accelerate

Down Arrow: Decelerate / brake

Speed gradually increases over time.

Collision with trees or rocks causes the skier to fall and ends the game.

Passing cleanly between a pair of flag gates awards +100 points.

Missing a flag gate deducts points.

Display HUD:

Top-left: Current Score, Speed (km/h), Distance traveled

Top-right: High Score (persisted in localStorage)

On game over, show final score and stats. Press Space to restart.

Audio

Generate wind sound using Web Audio API (oscillator + noise).

Wind volume and intensity should increase with the skier’s speed.

Additional Polish

Smooth camera follow with slight lag and tilt.

Reasonable lighting (ambient + directional + point lights from lamps).

Responsive to window resize.

Clear visual feedback on collision and scoring.

Generate the full, ready-to-run HTML file.

Local Deployment Unrestricted Version Qwen 3.8-Flash-Next (125B) Installation Tutorial | Strata