DavidAU Qwen3.8 27B Turbo Brilliance 50X-Reasoning... tested - Local LLM setup

Summary

Luke’s Dev Lab tested DavidAU’s Qwen 3.8 27B Turbo Brilliance 50X Reasoning Instruct NEO MAX MTP, using Q4_K_M quantization with MTP. The reviewer described it as a strong fine-tune, particularly for agentic coding tasks, where it often worked with little supervision.

Reasoning was configured to X-high, which may have affected the results and token usage.

Setup and performance

The model ran on Ubuntu Server with llama.cpp and was tested on an RTX 2000 and an RTX 5090. The reviewer also provided a configuration for a 16 GB GPU. The Q4_K_M setup can use system memory, with 32 GB of DDR4 RAM available for overflow.

• Generation speed was just under 11 tokens/sec on the RTX 2000 and about 118 tokens/sec on the RTX 5090.
• Prefill speeds varied substantially by GPU and whether the prompt was cached.
• On longer prefills, the RTX 5090 reached roughly 2,600 tokens/sec uncached and 144,000 tokens/sec cached.

Benchmark results

• 256K-context memory test: The model found the inserted information in 14 of 15 runs, missing it once at 25% context depth. The reviewer noted that performance looked strongest near the beginning and end of the context.
• Reasoning: 40/48 (83%) overall—100% on easy and medium questions, 92% on hard questions, and 42% on expert questions.
• Python challenges: 82/100. The model answered every challenge.

Coding and agentic tests

• Kanban web app: Produced a polished interface with working task labels, completion indicators, drag-and-drop, column movement, search, and a keyboard shortcut. The reviewer noted a small flicker when moving columns. No follow-up prompts were needed.
• Sand physics simulator: Most features worked, including material placement, acid effects, pause controls, and clearing the canvas. A known water-and-sand teleporting bug remained.
• Dungeon crawler: Passed the ray-casting, lighting, and dungeon-generation requirements.
• Blender asset: Created a detailed “starfall lantern” with appealing lighting, shadows, textures, and geometry. The reviewer prompted it once to provide missing screenshots.
• Godot 3D game: Completed a game with keys, moving platforms and hazards, falling, checkpoints, death and respawn, and a finish goal. The reviewer considered it functional and visually decent.

Across these tasks, the model generally required no additional prompting after the initial instruction, with few context compactions. The reviewer’s overall assessment was that it is a capable, hands-off fine-tune—not literally a “Pong to the Matrix” leap. Results could differ with a lower reasoning setting.

Sponsor segment

The video also features Egoite, a free Chromium browser for Windows and Mac designed for use by people and AI agents. The sponsor describes features such as importing existing Chrome sessions, running agents in isolated background spaces, and exposing browser actions through JavaScript functions. Its advertised speed and token-use improvements are sponsor claims, not results from the model tests.

Main speakers and sources

• Primary speaker/reviewer: Luke’s Dev Lab
• Model creator: DavidAU
• Sponsor: CitroLabs / Egoite