Gyro S 3.8 Flash Next tested - 16GB Local LLM setup

Summary

Luke’s Dev Lab tested Gyro S, a lower-memory variant based on Qwen 3.8 Flash Next, running locally through Strata as an alternative to LM CPP.

The target system was an RTX 2000 Ada with 16 GB of VRAM and 32 GB of DDR4 RAM, running Ubuntu Server. A 5090 was also present, but was used only for background benchmarks—not as the target hardware.

Performance and Benchmark Results

• 128K context performance: Uncached prompt processing started at around 400 tokens per second and reached roughly 950 tokens per second on larger prefills. Cached processing ranged from about 2,600 to 77,000 tokens per second on a 32K prefill. Generation speed was about 57 tokens per second.
• Memory recall: The model found planted data at five positions across a 128K context, passing all 15 trials.
• Reasoning: It scored 37/48 (about 77%) in the main results: all easy questions correct, 92% on medium, 83% on hard, and 33% on expert questions. The video’s closing recap gives a different overall percentage, so the detailed score is the clearest result.
• Python challenges: It answered all 100 challenges and got 97 correct.

Coding and Tool-Use Tests

• Seating planner: Produced a polished front-end app with table assignment and swapping, guest search and filters, light and dark modes, notifications, and drag-and-drop assignment. It worked without additional prompting, though the session used two compactions.
• Sand physics: The first version supported sand, water, acid, walls, and erasing, along with pause controls. It looked and ran well, but sand could incorrectly move upward through water. The second version added oil, lava, and fire; it worked well but required an extra prompt after a test process hung.
• Dungeon crawler: The first version’s ray-casting and map generation worked smoothly without additional prompting. Version two added a minimap, enemies, lives, and new dungeon generation. Movement could skip past intended positions, and the model needed an extra prompt after another test timeout.
• Blender via MCP: Created a detailed, lit lantern asset with materials and hanging elements. Blender crashed repeatedly while the model generated screenshots. The reviewer reports that the model eventually identified the screenshot-rendering and MCP timeout issue and adjusted its approach.
• Godot via MCP: Built a 3D platformer with keys, hazards, checkpoints, moving platforms, sprinting, and adjustable field of view. The reviewer liked the result, though the bloom was strong and a gap initially made the route impossible until two platforms were added. The test consumed many tokens because the model spent considerable effort trying to test the game.
• Remotion: Used to create an animated video recap of the results. The reviewer plans to incorporate this into future test videos.

Overall Assessment

Gyro S performed strongly on this 16 GB GPU setup, passing the project challenges and doing especially well on memory recall and Python. The main caveats were weaker expert reasoning, occasional test-process timeouts, and tool-use issues such as Blender crashes during screenshot generation.

Source

The main speaker and evaluator is Luke from Luke’s Dev Lab. The model is based on Qwen 3.8 Flash Next, and its model card is hosted on Hugging Face.