LAB-013/2026-10-24TESTING

Qwen3.8 Flash-Next 125B: Mac Studio M5 Ultra Coding Benchmark

I ran a single-shot coding test against the 125B Qwen3.8 Flash-Next at IQ3_S on a Mac Studio M5 Ultra with 96 GB of unified memory, served by Strata at 256K context and driven through two harnesses: the VS Code Copilot harness and OpenCode. The machine itself (GPU memory limits, quant choice, and headless serving) is documented in the Strata setup write-up.

01 / Result

Local inference snapshot

Same prompt and runtime per harness. OpenCode run metrics pending.

Hardware: M5 Ultra96 GB unified
Model: 125B MoEQwen3.8 Flash-Next ↗
Quant: IQ3_S3.5-bit GGUF ↗
Context: 256KStrata · Metal

02 / Configuration

Runtime settings

Full setup steps (SSH, GPU memory cap, Strata install, and network access) are in the Mac Studio M5 Ultra write-up.

Model: ISTA-DASLab/Qwen3.8-Flash-Next-GSQ-RCO-GGUF ↗
Quantization: IQ3_S · 3.5 bpw · ~63 GB resident
Runtime: Strata · Metal · MTP draft layer
Context: 256,000 tokens
Client: VS Code Copilot harness · OpenCode
Hardware: Mac Studio M5 Ultra · 96 GB unified memory
GPU memory cap: iogpu.wired_limit_mb=90112 · 88 GB wired
Access: LAN · OpenAI-compatible API on port 8080

03 / Showcase

Generated projects

Each response was saved as a standalone HTML file and is rendered below without modification.

Output 01 / VS Code Copilot harness / Three.js voxel art

Voxel pagoda garden

Tokens: 14.7M in / 152k out
Time: 24m 13s
Decode: ~130 t/s
Prefill: 1,700–2,500 t/s

Prompt

“Design and create a very creative, elaborate, and detailed voxel art scene of a pagoda in a beautiful garden with trees, including some cherry blossoms. Make the scene impressive and varied, and use colorful voxels. Use whatever libraries you need to get this done, but make sure it fits in a single HTML file. The settings should be hideable and able to be opened again. This should also work on mobile.”

Open Benchmark

Output 02 / OpenCode harness / Three.js voxel art

Voxel pagoda garden

Tokens: TBD
Time: TBD
Decode: TBD

Prompt

“Design and create a very creative, elaborate, and detailed voxel art scene of a pagoda in a beautiful garden with trees, including some cherry blossoms. Make the scene impressive and varied, and use colorful voxels. Use whatever libraries you need to get this done, but make sure it fits in a single HTML file. The settings should be hideable and able to be opened again. This should also work on mobile.”

Open Benchmark