Qwen3.8-Flash-Next Shrunk to 58GB — 50% Fewer Experts, 98.7% Coding Performance!

4 Просмотры

Спасибо! Поделитесь с друзьями!

Вам не понравилось видео. Спасибо за то что поделились своим мнением!

Издатель Oct 2, 2026

Qwen3.8-Flash-Next is a massive Mixture-of-Experts model, but this new GSQ-RCO Coder variant from ISTA-DASLab takes a very different approach to compression.

Instead of simply quantizing everything and hoping for the best, the model uses RCO — Riemannian Constrained Optimization — to selectively retain the experts that matter most for coding and agentic workloads, reducing the available expert pool from 512 experts to 256.
The remaining weights are then compressed using GSQ — Gumbel-Softmax Quantization, bringing the model down to around 58.4GB in GGUF form, with roughly 29.6GB for the transformer itself.

Chapters:
00:00 - Introduction
02:20 - Running QWEN3.8-FLASH-NEXT-GSQ-RCO in LM Studio
05:44 - Comparing QWEN3.8-GSQ-RCO vs QWEN3.8-Q4 for .NET Migration
08:12 - Comparing QWEN3.8-GSQ-RCO vs QWEN3.8-Q4 for SVG Image

My AI related Courses for you to empower today:
Learn my latest course launched this week on ZERO Code AI Tests: Automate Software Tests in Plain English: https://www.udemy.com/course/zero-code-ai-testing/?couponCode=LEARN_SEP_26

Use code: LEARN_SEP_26 for 85% OFF (limited time)

My Other AI related Courses for you to empower today:
To Learn more about building your own AI Agents and Testing AI Agents, follow the courses below available in Udemy with Discount coupon codes.

Use coupon code : LEARN_SEP_26